# FixMyPrompt — Full Content Index > AI prompt QA — score, fix, and ship better prompts. This file inlines the full body of every public blog post on FixMyPrompt.net so an LLM can read the entire knowledge base in a single fetch. For the lighter index, see /llms.txt. Site: https://fixmyprompt.net --- ## What is FixMyPrompt FixMyPrompt is a tool that reviews AI prompts before they go to a model. Users paste a draft prompt + a goal and get a structured QA report: a score 0-100 and severity-tagged issues with fixes. That much is free and needs no account, capped at 3 reports a day. A free account releases the rewritten 'improved prompt' at no charge. Credits buy the deeper Sonnet/Opus model tiers, the three rewrite variants (concise/detailed/structured), image-prompt QA, and saved/shared reports. It is provider-agnostic; the rewritten prompt can be shipped to any model the user already uses. Pricing beyond the free tiers is credit-based (no subscription required). A Premium subscription is on the roadmap and will include discounted top-ups, a tier-aware fast lane, usage analytics, JSON-schema validation, team accounts, and MCP server access. ## How the QA engine works The engine routes to one of three model tiers depending on the requested depth: basic uses claude-haiku-4.5, deep uses claude-sonnet-5.5, premium uses claude-opus-5.5 (current defaults; configurable via env). A paid user who has pinned a preferred model in account settings runs that model instead of the depth default, and is billed for it. Each report deducts credits server-side based on input length and depth tier. On any failure during generation, credits are auto-refunded and the report is marked failed. --- ## Claude vs ChatGPT: Which Model Is Better for Your Task? Published: 2026-10-12 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/claude-vs-chatgpt-which-is-better-for-your-task Neither Claude nor ChatGPT wins every task. Learn where each model excels, where it falls short, and how to adapt the same prompt for better results in both. You paste the same prompt into Claude and ChatGPT, and one of them gives you something usable while the other gives you something that sounds like a press release written by a committee. The frustrating part is that next week, on a different task, the results flip. Neither model is universally better. They're built differently, they weight different things, and a prompt that lands well in one can fall flat in the other for reasons that aren't obvious until you understand what each model is actually optimized for. Here's a practical breakdown of where each model tends to win, where it tends to lose, and how to adjust your prompts when you switch between them. ## Where ChatGPT tends to pull ahead ChatGPT is well-suited to structured tasks where the output needs to follow a predictable shape. Customer support scripts, technical documentation, step-by-step instructions, quick factual lookups: these are areas where its preference for organized, formatted output works in your favor. GPT-3.5 is optimized for speed, which makes it useful when you need volume fast and the task is relatively routine. GPT-4 adds more nuance and handles longer context better, but the underlying tendency toward structured output stays consistent across versions. If your workflow connects writing with research or analysis tools, ChatGPT's integrations give it a practical edge. It's built for rapid multi-format production in ways that make it easier to slot into a pipeline where the output feeds something else downstream. For step-by-step reasoning tasks, ChatGPT tends to be more reliable. If you ask it to walk through a problem methodically, it will usually do exactly that, in a format you can follow and check. ## Where Claude tends to pull ahead Claude's strengths cluster around language quality and context retention. For long-form writing, creative work, and anything that requires a consistent voice across many paragraphs, Claude tends to produce smoother, more human-feeling output. Its tone refinement capabilities let it adjust writing style and emotional register more naturally, which matters when you're editing something that needs to feel like a person wrote it. The context handling is worth paying attention to. Claude is particularly effective at working with long documents and retaining what was said earlier in a conversation. If you're synthesizing research into a narrative, or editing a long draft where earlier decisions need to stay consistent with later ones, Claude holds that thread better. Users have reported that Claude tends to understand what a prompt is asking with less back-and-forth clarification, which makes it a more efficient partner for editing tasks specifically. You spend less time re-explaining what you meant and more time working with the output. Claude Sonnet, a specialized version, is designed specifically for poetry and creative writing tasks, so if that's your use case, it's worth trying that variant rather than the default. ## The same prompt, two different results Here's where it gets practical. The same prompt, word for word, can produce noticeably different results depending on which model you're running it in. That's not a bug or a sign that one model is broken. It reflects genuine differences in what each model weights. When you're adapting a prompt across models, the most important thing is to run the same prompt word for word first. That gives you a real comparison rather than a comparison contaminated by your own rewrites. Once you see where the output diverges, you can make targeted adjustments. A few patterns that tend to hold: **For Claude**, you can often be less prescriptive about format. Claude tends to infer what you want from context, so a prompt that describes the goal and the audience will often get you further than one that specifies every structural element. If you want a particular tone, describe it in terms of how it should feel to the reader rather than listing adjectives. **For ChatGPT**, being explicit about structure usually pays off. If you want bullet points, say so. If you want a specific number of sections, name them. If you want the output to avoid a particular format, say that too. ChatGPT responds well to constraints because it's built to follow structured instructions precisely. **For critique and refinement tasks**, Claude's collaborative approach tends to produce better outcomes. If you're asking a model to push back on your draft, identify weaknesses, or suggest alternatives, Claude's tendency toward nuanced critique is more useful than a model that defaults to agreement. **For tasks that feed into other tools**, ChatGPT's structured output is easier to parse downstream. If the next step in your workflow is a script or a spreadsheet or a formatted document, ChatGPT's preference for clean structure saves you cleanup work. ## Switching models mid-project One situation that catches people off-guard is switching models partway through a project. You might start a research synthesis in ChatGPT, then move to Claude for the writing phase. The problem is that each model has its own sense of what the project is, and a prompt that worked in one context won't automatically carry its meaning into the other. When you switch, treat it as a fresh start. Bring the relevant context explicitly into your first prompt in the new model. Don't assume the new model knows what you've already established. Claude's strength at retaining context within a conversation doesn't help you if you're starting a new conversation without that context loaded in. Claude's agentic product, which integrates directly into the chat window and works on the user's file system, is worth knowing about if you're doing file-heavy work. It changes how you'd structure prompts for tasks that involve reading or writing files, since the model can interact with your file system directly rather than requiring you to paste content in. ## Choosing based on the task, not the brand The practical takeaway is to stop thinking about which model is better and start thinking about which model is better for this specific task. ChatGPT for structured outputs, technical specificity, and workflows that connect to other tools. Claude for long-form writing, creative work, editing, and anything that requires holding a lot of context at once. Your prompt is the other variable. A well-constructed prompt will get more out of either model than a vague one will, and a prompt that's calibrated for one model's tendencies will outperform the same prompt dropped unchanged into the other. If you're getting generic or off-target results and you're not sure whether the problem is the model or the prompt, the prompt is usually worth examining first. [FixMyPrompt](https://fixmyprompt.com/try) scores your prompt, flags the specific weak spots, and rewrites it, which gives you a cleaner variable to test across models. --- ## Sources The verified facts used in this post are drawn from the following sources. Because the original source URLs were not provided with the fact set, readers are encouraged to verify current model capabilities directly through Anthropic's documentation at [anthropic.com](https://www.anthropic.com) and OpenAI's documentation at [openai.com](https://www.openai.com), as model capabilities and product features change frequently. ## Sources - https://portkey.ai/blog/prompting-chatgpt-vs-claude - https://www.igmguru.com/blog/claude-vs-chatgpt - https://draganvaragic.substack.com/p/ai-models-hierarchy-4-types-llms - https://promptessor.com/compare/chatgpt-vs-claude - https://zapier.com/blog/claude-vs-chatgpt - https://www.reddit.com/r/PromptEngineering/comments/1ps6fsr/i_tested_the_same_prompt_across_chatgpt_claude --- ## How Do You Get ChatGPT to Write a Wedding Toast That Sounds Like You? Published: 2026-10-08 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/how-to-ask-chatgpt-to-write-a-wedding-toast Learn how to use voice samples, one specific story, and a copy-paste prompt to get ChatGPT to write a wedding toast that actually sounds like you. You have been asked to give a wedding toast. Maybe you said yes before you really thought about it, and now the wedding is six weeks away and you are staring at a blank page. You know the couple better than anyone. You have the stories. You just cannot seem to get the words out in a way that sounds like you and not like a greeting card. ChatGPT can help with this. But if you just ask it to "write a wedding toast," you will get something that sounds like every other wedding toast ever given: polished, warm, and completely hollow. This guide shows you how to give it enough to work with so the toast actually sounds like it came from you. --- ## First, how to get started If you have never used ChatGPT before, here is all you need to do. Go to chatgpt.com and make a free account. It takes about two minutes. Once you are in, you will see a box at the bottom of the screen. Type your message there and press Enter. That is it. You cannot break anything, and if the answer it gives you is not right, you can just tell it so and ask again. --- ## Why a plain request gives you a generic toast When you type "write a wedding toast for my best friend Sarah," ChatGPT has almost nothing to go on. It does not know how you talk. It does not know whether you are the kind of person who makes a room laugh or the kind who makes a room cry. So it guesses, and its guess is the average of every toast it has ever seen. The fix is to give it two things before you ask for anything: samples of your own writing, and the specific story you want to tell. --- ## Step one: give it samples of how you actually write This is the most important step. Find two or three things you have written that sound like you. A long text message to a close friend. An email where you were telling a funny story. A card you wrote for someone's birthday. Even a Facebook post works. They do not need to be polished. They need to be real. Copy those into ChatGPT and tell it: "Here are some things I have written. Read them and figure out how I write. Notice my sentence length, my tone, whether I use humor, how formal or casual I am." ChatGPT will pull out patterns from what you give it: things like whether your sentences tend to be short or long, whether you lean toward warmth or humor, how you start and end thoughts. The more samples you give it, the more it has to work with. --- ## Step two: pick one story and tell it plainly A great wedding toast is almost always built around one specific moment. The time you both got lost driving to a concert and ended up having the best night of the trip. The way she described him after their third date. The thing he did when her dad was in the hospital. Before you write anything, just tell ChatGPT the story in your own words, as if you were telling it to a friend on the phone. Messy is fine. Details are good. "We were at this dive bar in Nashville and she turned to me and said..." is much more useful than "they have always been there for each other." --- ## The copy-paste block to use Here is the full set of instructions to give ChatGPT. Copy this, and change the parts in [square brackets] to your own details. --- *I need you to write a wedding toast for me to deliver out loud. First, read these writing samples and learn how I write: [paste your writing samples here]. Now here is the story I want to tell: [describe the story in your own words]. The couple is [names], and I am the [best man / maid of honor / father of the bride / etc.]. The tone should be [warm and a little funny / mostly heartfelt / light and celebratory]. Write the toast in my voice, using short sentences and everyday words, the way I would actually say it out loud. Do not use formal or stiff language. Do not use phrases that sound like a computer wrote them. Include the story I told you. End with something genuine, not a cliché. After the draft, give me a confidence score from 0 to 100 for how closely it matches my writing style, and tell me what you would change to get that score higher.* --- Here is what that looks like filled in with realistic details: *I need you to write a wedding toast for me to deliver out loud. First, read these writing samples and learn how I write: [three text messages pasted here, including one where I'm telling a funny story about a road trip]. Now here is the story I want to tell: Marcus and I have been friends since college. The night he met Diane, he called me at midnight just to say "I think I met someone." He never does that. I want to tell that story and connect it to who he is as a person. The couple is Marcus and Diane, and I am the best man. The tone should be warm and a little funny. Write the toast in my voice, using short sentences and everyday words, the way I would actually say it out loud. Do not use formal or stiff language. Do not use phrases that sound like a computer wrote them. Include the story I told you. End with something genuine, not a cliché. After the draft, give me a confidence score from 0 to 100 for how closely it matches my writing style, and tell me what you would change to get that score higher.* --- ## What to do with the first draft Read it out loud. This matters. A toast is a spoken thing, and your ear will catch what your eye misses. If a sentence makes you stumble, shorten it. If a phrase does not sound like something you would say, change it. If the story is there but the feeling is off, tell ChatGPT exactly that: "This part sounds too formal. I would say it more like this: [your version]." Then ask it to rewrite from scratch using that feedback. You can go back and forth as many times as you need. Each round, it gets closer. Keep going until you would be comfortable reading it cold at the reception. One thing to watch for: phrases that feel smooth but hollow. Things like "from this day forward, your journey begins." If something sounds like it could be on a stock photo, cut it or ask ChatGPT to replace it with something specific to your story. --- ## A few small things that make a big difference Keep the toast between two and three minutes when read aloud. That is roughly 300 to 450 words. Shorter is almost always better. Pick one story, not four. Four stories means none of them land. End on the couple, not on yourself. The last thing the room hears should be about them. --- ## One last thing If you want a second opinion on the instructions you put together before you hand them to ChatGPT, fixmyprompt.net/try will check them and suggest improvements for free, up to three times a day. Now go write something that will make them cry in the good way. --- ## Sources - OpenAI Developer Community: https://community.openai.com - Ruben Hassid: https://www.rubenhassid.com - Andrew Bolis: https://andrewbolis.com - Medium: https://medium.com ## Sources - https://medium.com/@syedhidayat901/5-prompts-that-make-chatgpt-sound-exactly-like-you-not-like-chatgpt-4646f15f16cf - https://www.linkedin.com/posts/ruben-hassid_how-to-prompt-chatgpt-to-write-exactly-activity-7352291141414748160-ldib - https://www.linkedin.com/posts/andrewbolis_copy-and-paste-this-prompt-to-make-chatgpt-activity-7389260625169920000-Bnst - https://community.openai.com/t/ask-for-help-what-is-the-ideal-prompt-to-make-the-content-more-humanized/1124343 - https://www.youtube.com/watch?v=11ux9v4luK4 - https://medium.com/@dutttarush9360/i-tried-125-chatgpt-prompts-heres-the-only-one-that-sounds-human-074d45fe252e --- ## How to Ask ChatGPT to Plan a Week of Meals for Your Family Published: 2026-10-01 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/how-to-ask-chatgpt-to-plan-weekly-meals Learn how to give ChatGPT the right details about your budget, schedule, and picky eaters to get a weekly meal plan you will actually follow. If you have ever searched "healthy meal plan" and printed something out, only to abandon it by Wednesday, you are not doing anything wrong. The plan was just built for nobody in particular. ChatGPT can do something different. You can tell it about your actual family, your actual budget, and the fact that nobody in your house will eat mushrooms, and it will build around all of that. This post shows you exactly how. --- ## First, how to get started Go to [chatgpt.com](https://chatgpt.com) and make a free account. Once you are in, you will see a box at the bottom of the screen. Type your question or instructions there and press Enter. That is the whole thing. You cannot break anything, and if the answer it gives you is not what you wanted, you can just tell it that and ask again. --- ## Why generic meal plans fall apart When you ask for "healthy meals for the week," you get a plan designed for a fictional average person. It assumes you have an hour to cook on a Tuesday. It does not know you have a picky eight-year-old or that you already have half a bag of lentils in the cupboard. It does not know your grocery store, your budget, or that you hate doing dishes. So you look at Day 3 and think, "I am not making that," and the whole thing goes in the recycling bin. The more specific your instructions, the more useful the answer. A plan built around your real life is one you might actually follow. --- ## What to include when you ask for a meal plan Here is what makes the difference between a plan you use and one you ignore. **Your family size and who is eating.** Four people eating dinner is different from two adults and two small children. Portion sizes change, and so does the grocery list. **Any dietary restrictions.** Be specific. "We avoid pork" or "my daughter is lactose intolerant" gives it something to work with. One important note: ChatGPT cannot guarantee that any recipe is safe for a serious allergy. If someone in your family has a severe allergy, check every single ingredient yourself before cooking. **How much time you actually have.** You can ask for meals where the active cooking time is under 30 minutes. That is a real thing you can specify, and it will stick to it. **Your weekly grocery budget.** A family of four might set a budget of $150 for the week, covering breakfasts, lunches, snacks, and dinners. A single adult might work with around $60, focusing on whole foods. Put your real number in. **How many nights you will actually cook.** If you know you order takeout on Fridays and eat at your sister's on Sundays, say so. Asking it to plan five cooking nights instead of seven means the plan fits your week, not an imaginary one. **What equipment you have.** If you mention an Instant Pot or an air fryer, it can lean on those. More usefully, if you do not have certain equipment, say so, and it will stop suggesting recipes that need it. **What cuisines your family enjoys.** If your household loves Mexican food and Asian dishes, say that. You will get a plan that feels familiar rather than one that feels like a homework assignment. **What is already in your pantry.** If you have rice, canned tomatoes, and chicken thighs, tell it. It can build meals around what you already have, which cuts your grocery bill and reduces waste. --- ## The copy-paste block Here is a set of instructions you can copy and use. Change everything in [square brackets] to match your own situation. ``` Please create a 7-day meal plan for [family size, e.g. a family of four: two adults and two children aged 8 and 11]. Dietary needs: [list any restrictions, e.g. no shellfish, one person is lactose intolerant]. We will cook on [number of nights, e.g. five nights]. The other nights we have leftovers or eat out. Active cooking time should be under [time, e.g. 30 minutes] per meal. Weekly grocery budget: [your budget, e.g. $120], covering dinners and lunches. We enjoy [cuisines or food types, e.g. Mexican and Italian food]. We do not like [dislikes, e.g. mushrooms or very spicy food]. Equipment we have: [e.g. a standard oven, stovetop, and slow cooker]. We already have in the pantry: [e.g. rice, pasta, olive oil, canned beans, chicken stock]. Please include a consolidated grocery list with estimated prices and notes on where a generic brand would work just as well as a name brand. You can also include approximate calories and protein per dinner if that is easy to add. ``` Here is what that looks like filled in with real details: ``` Please create a 7-day meal plan for a family of four: two adults and two children aged 9 and 13. Dietary needs: no pork, and one adult is vegetarian so please make dinners that work for everyone or note easy swaps. We will cook on five nights. The other nights we have leftovers or order pizza. Active cooking time should be under 30 minutes per meal. Weekly grocery budget: $150, covering dinners and packed lunches for the kids. We enjoy Mexican and Asian food. We do not like mushrooms or very fishy-tasting fish. Equipment we have: a standard oven, stovetop, and an air fryer. We already have in the pantry: rice, canned black beans, soy sauce, olive oil, and a bag of frozen corn. Please include a consolidated grocery list with estimated prices and notes on where a generic brand would work just as well as a name brand. ``` Copy the first block, fill it in with your details, paste it into the box at chatgpt.com, and press Enter. --- ## What you will get back ChatGPT will give you a day-by-day plan, a single grocery list that covers the whole week, and notes on where you can save money by choosing store brands. If you asked for nutritional information, it will include rough calorie and protein counts per meal. It pulls all your constraints together so you are not cross-referencing five different things. If something in the plan does not work, just say so. "Can you swap Wednesday's dinner for something that uses the leftover chicken?" works perfectly well. You are having a conversation, not filling out a form. --- ## A few things worth knowing The plan it gives you is a starting point. Prices will vary by where you live and what is on sale that week. Treat the grocery estimates as a rough guide, not a guarantee. And again: if anyone in your family has a serious food allergy, read every recipe ingredient yourself. Do not rely on ChatGPT to catch every hidden source of an allergen. It sometimes makes things up or misses details, and with allergies, that matters. --- ## One more thing If you want to check whether your instructions are as clear and specific as they could be before you use them, [fixmyprompt.net/try](https://fixmyprompt.net/try) will look them over for free, up to three times a day. --- ## Sources - Little Potatoes: https://www.littlepotatoes.com - Coursiv Blog: https://www.coursiv.io/blog - MealThinker: https://www.mealthinker.com - AI Academy: https://www.aiacademy.com - Chris Kresser: https://chriskresser.com ## Sources - https://mealthinker.com/blog/chatgpt-meal-planning-prompts - https://coursiv.io/blog/ai-meal-planner - https://academy.techpresso.co/prompts/chatgpt-prompts-meal-planning - https://www.littlepotatoes.com/blog/meal-planning-with-ai - https://chriskresser.com/turn-chatgpt-into-your-personal-meal-planner - https://www.eatforendurance.com/post/chatgpt-meal-planning --- ## Why Are ChatGPT's Answers So Short Now? Published: 2026-09-28 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/why-are-chatgpt-answers-so-short GPT-5.6 defaults to shorter, more cautious replies than earlier versions. Learn what changed, why it happens, and how to prompt for the depth you actually need. If you've been using ChatGPT for a while, you've probably noticed something feels off lately. You ask a question that used to get you three detailed paragraphs, and now you get four sentences and a shrug. You're not imagining it, and you're not alone. Here's what actually changed with GPT-5.6, why the model behaves the way it does, and what you can type right now to get the depth you're looking for. --- ## What Actually Changed After GPT-5.6 The shift is real. GPT-4 was known for multi-paragraph answers that worked through a problem before landing on a conclusion. GPT-5.6 defaults to shorter, more compressed responses, and users have noticed the difference. Two other patterns have emerged alongside the terseness. Hedging has increased, with phrases like "it is important to note" and "it depends" appearing more often than they did in earlier versions. And refusals on perfectly benign queries have become more common, suggesting the model is now calibrated to err on the side of caution rather than completeness. There's also a transparency issue. Earlier versions would show how long the model spent reasoning before answering. Some users report that GPT-5.6 sometimes skips displaying that thinking time entirely, making it harder to tell whether the model worked through a problem or just pattern-matched to a quick reply. The result is a model that can feel less useful for anything requiring nuance, even when the underlying capability is still there. --- ## Why the Model Behaves This Way A few mechanics are worth understanding. **Default behavior favors brevity.** The model's out-of-the-box behavior now leans toward short answers. If you don't tell it otherwise, it will give you the compressed version. This isn't a bug so much as a calibration choice, but it means the burden of specifying what you want has shifted to you. **Reasoning effort is adjustable.** GPT-5.6 offers processing options, Medium, High, and Extra High, that affect how much reasoning the model applies before answering. A quick reply on Medium effort and a careful reply on Extra High can look very different for the same question. Not every question needs Extra High, but for anything complex, the default may be undershooting. **Long outputs have a reliability ceiling.** User experience suggests that answers longer than roughly 500 words can start to lose coherence or trail off. This may be part of why the model defaults to shorter responses. Knowing this helps you work with it rather than against it: ask for depth in sections rather than demanding one enormous answer. **Context doesn't persist across sessions.** The model holds your conversation in memory as long as the current page stays open. Refresh or close it, and that context is gone. If you've been building on a thread and then come back to a new session, the model has no memory of what you established before, which can make follow-up answers feel shallow or disconnected. --- ## How to Ask for the Depth You Actually Want The good news is that small prompt changes produce meaningfully different outputs. Here's what works. **State the length explicitly.** If you want a thorough answer, say so. "Give me a detailed explanation, at least 400 words" or "walk me through this step by step" signals that brevity is not the goal. The model's default favors short, so you have to override it. **Specify the complexity level.** Adding "explain this at a level intended for someone with a PhD in the field" or "explain this as you would to a high school student" does more than adjust vocabulary. It signals the depth of reasoning you expect. A PhD-level explanation implies you want the nuances, the caveats, and the edge cases, not just the summary. **Set the reasoning effort before you ask.** If your interface gives you the option to select Medium, High, or Extra High processing, use it deliberately. For a quick factual lookup, Medium is fine. For anything where you need the model to actually think through tradeoffs, set it higher before you send the prompt. **Break long requests into sections.** Rather than asking for a 1,000-word analysis in one shot, ask for the first part, confirm it's going in the right direction, then ask for the next. This sidesteps the coherence issues that can appear in very long single outputs. **Ask for structure when you need it.** GPT-5.6 can generate structured documents, spreadsheets, and presentations, not just plain text. If you need a financial model or a formatted report rather than a wall of prose, ask for that format explicitly. The model won't default to it. **Push back on hedging.** If the model gives you an "it depends" answer when you need a concrete recommendation, follow up with "given those variables, what would you actually recommend?" or "assume the most common scenario and give me a direct answer." The model is capable of being more definitive; it just needs permission. --- ## The Underlying Problem Is Usually the Prompt Here's the pattern that shows up again and again: the model has the capability to give a thorough, useful answer, but the prompt didn't ask for one clearly enough. GPT-5.6 is more conservative by default, so a vague or underspecified prompt gets a vague, underspecified reply. This is exactly the problem FixMyPrompt was built to solve. You paste your prompt, get a score from 0 to 100 and see every weak spot flagged with a specific fix, free, no signup. A free account adds the rewritten version, ready to use. It works with ChatGPT, Claude, Gemini, and image models. If your prompts are getting short or generic answers, the score usually tells you why. [Try it at /try.](https://fixmyprompt.com/try) The terseness of GPT-5.6 is real, but it's also workable. The model still has depth. You just have to ask for it in a way that makes clear you want it. --- ## Sources The claims in this post are based on the verified facts provided for this article. No external URLs were supplied as sources, so no external links are listed here. ## Sources - https://fixmyprompt.net/blogs/19-is-chatgpt-getting-worse-2026 - https://www.zdnet.com/article/10-chatgpt-prompt-tricks-i-use-to-get-the-best-results-faster - https://ttms.com/chatgpt-5-6-in-practice-initial-compliments-and-disappointments - https://blog.risingstack.com/gpt-5-6-sol-terra-luna-chatgpt-changes - https://news.ycombinator.com/item?id=49199357 - https://community.openai.com/t/gpt-5-6-sol-high-reasoning-appears-broken-instant-replies-no-deep-thinking-file-workflows-failing/1394605 --- ## How to Ask ChatGPT to Rewrite Your Resume for a Specific Job Published: 2026-09-24 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/how-to-ask-chatgpt-to-rewrite-your-resume Learn the exact prompt to use, why pasting the full job posting matters, and what to check before you send your ChatGPT-rewritten resume anywhere. You spent an hour on your resume. You tailored it, you thought, for the job you really want. Then you heard nothing. The problem is usually not your experience. It is that a resume written for everyone tends to speak to no one. A job posting tells you exactly what a company wants to hear. This post shows you how to use ChatGPT to close that gap, step by step, even if you have never opened ChatGPT before. ## Getting started with ChatGPT Go to chatgpt.com and make a free account. It takes about two minutes. Once you are in, you will see a box at the bottom of the screen. Type there and press enter. That is all there is to it. You cannot break anything, and if the answer it gives you is not what you wanted, just say so and ask again. ## Why a generic resume tends to fall flat A resume that lists your experience without connecting it to a specific job makes the reader do all the work. They have to figure out whether you are a fit. Most will not bother. When you customize your resume for one role, you are doing that work for them, showing them the match rather than asking them to find it. ChatGPT can help you make that match faster. The key is giving it enough to work with. ## What to gather before you start You need two things open on your screen before you type anything. First, your current resume. Copy the whole text of it. Second, the full job posting for the role you want. Copy that too, the whole thing, including the requirements section that lists every skill and qualification. The reason this matters: when ChatGPT can see the exact words the company used, it can weave those same words into your resume. Recruiters and the software companies use to screen applications both look for that specific language. Pasting the posting in is not a shortcut. It is the whole point. ## The instructions to copy and use Here is what to type into ChatGPT. Copy this block, then swap out everything in square brackets for your own details. --- *Rewrite my resume for the [Job Title] role at [Company] using this job description: [Paste the full job description here]. Here is my current resume: [Paste your full resume here]. Keep everything accurate to my real experience. Turn duties into achievements where you can. Format it so it is clean and easy to read. Do not add anything I have not done or claimed.* --- Here is what that looks like filled in: *Rewrite my resume for the Operations Coordinator role at Maple Street Health using this job description: "We are looking for an organized, detail-oriented coordinator to manage scheduling, vendor relationships, and internal communications across a 40-person team..." [rest of posting]. Here is my current resume: "Jane Doe, Operations Coordinator, 2018 to present, managed daily scheduling for a team of 35..." [rest of resume]. Keep everything accurate to my real experience. Turn duties into achievements where you can. Format it so it is clean and easy to read. Do not add anything I have not done or claimed.* --- The more detail you give, the more useful the answer will be. Include your job titles, the years you worked each role, any numbers you remember (team sizes, budgets, results), and any skills you want to make sure appear. The fuller the picture you give, the better the rewrite will fit you. ## What ChatGPT will do with this It will look at the language in the job posting and find the places in your experience where those things show up. Then it will rewrite your bullet points to reflect that connection. If your old resume said "responsible for scheduling," it might come back as "coordinated scheduling for a 35-person team, reducing conflicts by consolidating requests into a single weekly system." That shift, from a duty to a result, is what makes a resume feel like evidence rather than a list. You can also ask it to focus on one section at a time. If your summary paragraph feels weak, ask it to rewrite just that. If you want your most recent job to carry more weight, ask it to work on that section specifically. ## Before you send it anywhere This is the part people skip, and it is the most important part. Read every line of the rewritten resume out loud. Check every employer name, every job title, every date, every skill, every number, and every achievement. ChatGPT sometimes makes things up. It might round a number up, invent a responsibility you never had, or phrase something in a way that sounds like a claim you cannot back up in an interview. You are the only one who knows what is true. Go through it the way you would proofread a legal document, slowly, looking for anything that does not match your real history. If something is wrong, fix it yourself or ask ChatGPT to correct it. Tell it exactly what the problem is: "You wrote that I managed a budget of $2 million. I never had budget responsibility. Please remove that." Also look at the format. Some companies use software to scan resumes before a human ever reads them. That software works best with plain, simple layouts. Avoid tables, columns, text boxes, and unusual fonts. A single-column document with clear headings is the safest choice. ## A few things worth keeping in mind The goal of all this is to show a real match between your real experience and what the company actually needs. It is not to make yourself sound like someone you are not. A resume that overstates your background might get you an interview you are not prepared for, which helps no one. ChatGPT is a writing tool. It can help you say what you have done more clearly and in language that fits the job. The substance still has to come from you. If the first version it gives you does not feel right, tell it that. Say "this sounds too formal" or "I want the tone to be warmer" or "can you make the bullet points shorter." It will try again. Keep going until it sounds like you, just a sharper version of you. ## One last thing If you would like someone, or something, to check whether your instructions to ChatGPT are as clear and complete as they could be, fixmyprompt.net/try does that free, three times a day. --- ## Sources - Nils Davis: https://nilsdavis.com/chatgpt-prompts-for-resume/ - Coursera: https://www.coursera.org/articles/chatgpt-resume - Teal: https://www.tealhq.com/post/chatgpt-prompts-for-resume - CandyCV: https://www.candycv.com/blog/chatgpt-prompts-for-resume/ ## Sources - https://www.linkedin.com/posts/nilsdavis_dont-ask-chatgpt-to-revise-my-resume-so-activity-7391905764145127424-nlaK - https://www.coursera.org/articles/chatgpt-resume - https://www.tealhq.com/post/great-chatgpt-prompts-for-your-resume - https://www.linkedin.com/posts/mikeo13rien_tailor-your-resume-for-every-job-you-apply-activity-7379905494716547074-rDFa - https://www.facebook.com/groups/officialartificialintelligence/posts/2501319813677778 - https://www.candycv.com/how-to/best-chatgpt-prompts-to-improve-your-resume-3-copy-paste-prompts-20 --- ## Why Does ChatGPT Keep Asking Clarifying Questions? Published: 2026-09-21 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/why-does-chatgpt-keep-asking-questions ChatGPT asks clarifying questions when your prompt leaves too many gaps. Learn why the loop happens and how to preempt it with better upfront prompts. You asked ChatGPT a question. Instead of answering, it shot back two or three questions of its own. You answered those. It asked another. By the time you got a response, you'd spent more time in a back-and-forth than it would have taken to just write the thing yourself. This isn't a bug or a quirk of your account. It's a predictable behavior with a clear cause, and once you understand it, you can stop it before it starts. --- ## Why ChatGPT Asks Instead of Answers ChatGPT is designed to ask clarifying questions to ensure it understands a query fully before responding. That sounds reasonable on paper. The problem is that the model doesn't always recognize when it actually lacks enough information to give a good answer. It may proceed confidently on a vague prompt, or it may pause and interrogate you on a prompt that was perfectly clear. The behavior comes from how the model predicts output. It's weighing possible interpretations of your request, and when the probability mass is spread across too many of them, asking a question is the path of least resistance. The model isn't reading your mind. It's hedging. This creates a loop that feels productive but often isn't. You answer the clarifying question, the model incorporates that, then surfaces another gap it could have identified from the start. Each exchange adds latency without adding much to the final output. --- ## The Real Cost of the Loop The questioning loop isn't just annoying. It degrades the quality of the conversation in a few ways. First, your answers to clarifying questions get appended to the chat as follow-up messages. That context is less structured than if you'd included it in the original prompt. The model treats a well-organized initial prompt differently from a scattered thread of corrections and additions. Second, the loop trains you to give less upfront. If you expect to be asked, you stop front-loading detail. That's the opposite of what produces good output. Third, and most practically: when you answer a clarifying question in the chat rather than editing your original prompt, you're building on a shaky foundation. The recommendation from people who study this closely is to avoid answering clarifying questions in the chat at all. Instead, use the Edit Prompt feature to add those details directly into your initial message and resubmit. You get a cleaner context window and a better response. --- ## When the Questions Are Actually Useful There's one mode where you should answer the questions rather than preempt them: Deep Research. In that mode, ChatGPT will automatically ask clarifying questions before it begins a longer research task, and those questions are doing real work. The model is scoping a multi-step process, not hedging on a simple request. Answer those in the chat. For standard conversations, though, the clarifying-questions step is something you want to handle yourself, before you hit send. --- ## How to Preempt the Loop The fix is front-loading. A prompt that leaves no obvious gaps gives the model less reason to pause and ask. Here's what that looks like in practice. **Specify your role and the reader's role.** "Write a cold email" is a gap-filled prompt. "Write a cold email from a freelance UX designer to a mid-size SaaS company's head of product, pitching a two-week audit" is not. The model knows who's speaking, who's being addressed, and what the goal is. **Include the constraint you'd otherwise be asked about.** Length, tone, format, audience, deadline context, whatever the obvious follow-up question would be. If you can predict the clarifying question, you can answer it preemptively in the prompt. **Add the context the model can't infer.** ChatGPT may provide outdated or incomplete information unless you include specific context. If your question depends on a recent event, a personal situation, or a niche domain, that information needs to be in the prompt. The model doesn't know your circumstances unless you tell it. **Tell it what you don't want.** Constraints are as useful as instructions. "Don't ask me clarifying questions, make reasonable assumptions and note them at the end" is a legitimate instruction that works. A before-and-after makes this concrete: *Before:* "Help me write a performance review for someone on my team." *After:* "Write a performance review for a mid-level software engineer who consistently ships on time but struggles with written communication. Tone should be direct and constructive. Length: 200-250 words. Assume the review will be read by both the employee and their skip-level manager." The second prompt leaves almost nothing to ask about. The model can go straight to the answer. --- ## A Structural Trick for Complex Requests If you're working with the API or building a more complex workflow, there's a useful pattern: separate the clarification phase from the generation phase explicitly. Run a first pass where the model's only job is to identify what it would need to know to answer well, then incorporate those answers into a second prompt that does the actual work. This keeps the two concerns from bleeding into each other and gives you cleaner outputs at each stage. For most people using ChatGPT in a standard chat window, the simpler version of this is just: write the prompt, ask yourself what the obvious follow-up question would be, answer it in the prompt, then send. --- ## The Underlying Pattern Every clarifying question ChatGPT asks is a signal. It's pointing at a gap in the prompt. That gap was always there. The question just made it visible. The more useful habit is to treat that signal as feedback on your prompting rather than as a conversation to have. When ChatGPT asks "what tone are you going for?" the right response isn't to type the answer into the chat. It's to note that tone is something you should specify upfront every time you ask for written content. Over time, this builds a mental checklist. Role, audience, format, length, constraints, context. Prompts that cover those bases get answers, not questions. --- A vague prompt is almost always the root cause of bad or generic output, whether that's a questioning loop, a response that misses the point, or an answer that's technically correct but useless for your situation. FixMyPrompt scores your prompt from 0 to 100, flags every weak spot with a specific fix, and rewrites it for you. If you've been stuck in the clarifying-questions loop, [paste your prompt and see what's missing](/try). --- ## Sources - How to Improve Your ChatGPT Prompts: https://www.pcmag.com/how-to/how-to-improve-your-chatgpt-prompts - How to get GPT to ask more questions? OpenAI Developer Community: https://community.openai.com/t/how-to-get-gpt-to-ask-more-questions/549185 - How to use ChatGPT effectively: Don't ask it questions: https://www.zdnet.com/article/how-to-use-chatgpt-effectively/ - The one prompt I paste at the start of every ChatGPT chat: https://www.zdnet.com/article/the-one-prompt-i-paste-at-the-start-of-every-chatgpt-chat/ ## Sources - https://publish.obsidian.md/petermilovcik/1+Notes/How+to+Improve+Your+ChatGPT+Prompts - https://community.openai.com/t/how-to-get-gpt-to-ask-more-questions-system-message/726656 - https://www.linkedin.com/posts/nicoleleffer_dont-ask-chatgpt-to-ask-you-clarifying-questions-activity-7314038063519707136-w5KH - https://www.reddit.com/r/ChatGPTPromptGenius/comments/1ukeqao/the_one_prompt_i_paste_at_the_start_of_every - https://medium.com/@lca199/why-cant-i-stop-answering-chatgpt-s-questions-15afeed08ecb - https://finance.yahoo.com/news/student-loan-debt-erased-asking-153756830.html --- ## How Do You Ask ChatGPT to Write a Eulogy? Published: 2026-09-17 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/how-to-ask-chatgpt-to-write-a-eulogy Learn how to prompt ChatGPT to write a eulogy that actually captures the person. What details to gather first and what to never let the AI invent. You have a funeral to help plan, and someone has asked you to speak. Or maybe you volunteered, because you loved this person and you want to do right by them. Either way, you are sitting with grief and a blank page, and the words are not coming. ChatGPT can help with this. It will not replace what you know about this person, but it can take the raw material you give it and shape it into something you can actually say out loud. The catch is that you have to give it enough to work with. Hand it nothing, and it will hand you back a speech about a kind soul who loved their family, which could describe almost anyone. Hand it the real stuff, and it can give you a real starting point. If you have never used ChatGPT before, here is how to get there. Go to chatgpt.com and make a free account. Once you are in, you will see a box at the bottom of the screen. Type your message there and press Enter. That is all there is to it. You cannot break anything, and if the answer it gives you is not right, you can just tell it that and ask again. ## Gather the real details before you type a single word This is the step most people skip, and it is the most important one. Writing a eulogy is hard when you are grieving, but if you sit down with a few questions first and just write whatever comes, you will have the raw material you need. The actual writing gets easier after that. Ask yourself things like: Where did they grow up, and what shaped them early on? What is one moment that shows exactly who they were? What did they do that nobody else did quite the same way? What did they care about most? What would make the people in that room smile and nod? One question worth sitting with: what is one memory that captures who they were? Not a summary of their whole life. One specific moment. The answer to that question is often the heart of the whole speech. Also think about the length you need. A main speaker usually aims for five to seven minutes. If several people are speaking, three to four minutes each works better. A five-minute speech is roughly 600 to 700 words, so you are not writing a novel. ## What to put in your instructions to ChatGPT The more specific you are, the better the answer will be. A good eulogy has a shape: it opens with something that immediately captures the person, moves through a couple of stories that show who they were to different people, touches on their humor or a quirk that made them them, and closes with something the room will carry home. You can ask ChatGPT to follow that shape. Here is a block of instructions you can copy and use. Copy this exactly, then change everything in [square brackets] to your own details. --- *Please help me write a eulogy for [name]. They were my [relationship, e.g. grandmother, best friend, father]. The service is [date or timeframe], and I will be speaking for about [length, e.g. five minutes].* *Here is what I want you to know about them:* *- They grew up [brief detail about their background or hometown]* *- One memory that captures who they were: [describe a specific moment or story]* *- Something they did that nobody else did quite the same way: [a habit, a phrase, a skill, a quirk]* *- What they meant to other people: [how friends, family, or community saw them]* *- Something that would make the room smile: [a funny or warm detail]* *- What I most want people to leave feeling: [e.g. grateful, comforted, like they knew her better]* *Please write a eulogy that opens with a specific moment or line that captures them right away. Include two or three short stories that show who they were. Leave room for [any topic to avoid, e.g. the circumstances of their death]. Use plain, warm language. Do not invent any details I have not given you.* --- Here is what that looks like filled in with real details: *Please help me write a eulogy for Margaret. She was my grandmother. The service is this Saturday, and I will be speaking for about five minutes.* *Here is what I want you to know about her:* *- She grew up on a dairy farm in rural Vermont, one of six kids, and never lost the habit of waking before sunrise.* *- One memory that captures who she was: every Sunday she made pie from scratch, and she always let me crimp the edges, even when I did it wrong. She never once took over.* *- Something she did that nobody else did quite the same way: she kept a small notebook in her apron pocket and wrote down things people said that made her laugh.* *- What she meant to other people: her neighbors called her first when something went wrong. She was the person you called.* *- Something that would make the room smile: she once told my cousin that the secret to a good marriage was "separate bathrooms and the same politics."* *- What I most want people to leave feeling: like they just spent a few more minutes with her.* *Please write a eulogy that opens with a specific moment or line that captures her right away. Include two or three short stories that show who she was. Use plain, warm language. Do not invent any details I have not given you.* ## What to do with the draft it gives you Read the whole thing through once before you change anything. Just read it for the shape, the feeling, whether it sounds like it is about her or about a stranger. That first read will tell you what is working and what needs your hand on it. Then go sentence by sentence. Any line that could describe anyone, replace it with something specific. Any word that does not sound like you, change it. ChatGPT does not know how you talk, so the voice will need adjusting. That is normal and expected. One thing to be careful about: ChatGPT sometimes fills in details it was not given. Read every sentence and ask yourself, did I tell it this? If the answer is no, take it out or replace it with something true. The people in that room knew this person. They will notice. If the eulogy touches on a death by suicide, keep the focus on the person's life, not the circumstances of their death. You can tell ChatGPT this directly in your instructions. ## A last thought You are not trying to write a perfect speech. You are trying to give the people in that room a few minutes with someone they loved. The details you gathered, the specific memories, the small true things, those are what will do that. ChatGPT can help you shape them. The material itself only you can provide. If you would like your instructions checked and improved before you use them, fixmyprompt.net/try does it free, three times a day. ## Sources - https://aiacademy.com/how-to-use-chatgpt-to-write-a-eulogy/ - https://vegavid.com/blog/chatgpt-prompts-for-eulogy/ - https://wheelchairjohn.com/chatgpt-eulogy/ - https://heatherstang.com/chatgpt-eulogy/ - https://www.businessinsider.com/how-to-write-eulogy-chatgpt-ai ## Sources - https://academy.techpresso.co/prompts/chatgpt-prompts-eulogies - https://vegavid.com/blog/how-to-use-ai-to-write-a-eulogy - https://wheelchairjohn.com/2026/07/30/using-chatgpt-to-write-a-eulogy - https://heatherstang.com/eulogy-writing-prompts - https://www.businessinsider.com/how-to-use-get-better-chatgpt-ai-prompt-guide - https://www.reddit.com/r/ChatGPTPromptGenius/comments/1dwu0t5/eulogy_writing --- ## Why Does ChatGPT Give a Different Answer Every Time? Published: 2026-09-10 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/chatgpt-different-answer-every-time ChatGPT generates responses fresh each time using probability, not stored answers. Learn how temperature, context, and prompt structure affect consistency. You asked ChatGPT the same question twice and got two completely different answers. Maybe the tone shifted, the structure changed, or it contradicted something it told you an hour ago. You're not imagining it, and your account isn't glitching. This is how the model works, and once you understand the mechanism, you can write prompts that fight back against it. --- ## ChatGPT doesn't retrieve answers, it generates them A search engine looks up a stored result. ChatGPT does something fundamentally different: it predicts the most likely next word, then the next, then the next, building a response from scratch every single time. There's no internal database of correct answers to pull from. The model generates text based on learned patterns, which means natural variation is baked in from the start. This is why two runs of the same prompt can produce responses that feel like they came from different writers. They kind of did, in the sense that each generation is a fresh probabilistic event, not a playback. --- ## Temperature is the main dial, but it's not the whole story If you've poked around the API or read anything about language models, you've probably seen the word "temperature." It controls how much randomness the model introduces when choosing between possible next words. Low settings (roughly 0 to 0.3) push the model toward its most confident predictions, producing tighter, more predictable output. Medium settings (0.4 to 0.7) let the model roam a little, which tends to produce responses that feel more natural and less robotic. High settings (0.8 to 1.0) open the door to more creative and varied responses, which is useful for brainstorming but rough for anything where you need the same answer twice. Setting temperature to 0 is the closest thing to a consistency switch, but it doesn't guarantee identical output. One documented case found that a single prompt run 1,000 times at temperature 0 still produced 80 distinct answers. The model's architecture introduces some variability that temperature alone can't eliminate. A related parameter, top_p, works alongside temperature to constrain the pool of words the model considers. Reducing both together gives you the best shot at stable output. --- ## Why "identical" prompts aren't actually identical to the model Here's something that trips people up: you can type the exact same sentence twice and the model may not treat it the same way. Context is cumulative. If you asked three questions before this one, those questions are part of the input the model sees. The surrounding conversation steers the answer, so two prompts that look identical to you may arrive at the model wrapped in completely different context. The wording of a prompt and the prior conversation context can both shift the output significantly, even when the user's typed input appears unchanged. There's also the memory problem in long conversations. ChatGPT processes text in chunks and doesn't perfectly retain everything said earlier in a session. This can produce contradictions within a single conversation, where the model gives you one answer early on and a different one later, not because it changed its mind, but because it's working from an incomplete picture of what was said. --- ## The model itself changes underneath you Even if you control for temperature and context, the model version can change over time. Providers update and swap models regularly, which affects how responses are generated. A prompt that reliably produced a specific format last month may behave differently today, not because your prompt changed, but because the model did. This is worth knowing if you're building anything that depends on consistent output, whether that's a workflow, a template, or a repeatable research process. --- ## What consistent output actually looks like in practice A WSJ-affiliated study found ChatGPT gave consistent answers only 73% of the time across 10 identical prompts. That's not a failure mode, it's the expected behavior of a probabilistic system. The question is what you do about it. A few approaches that actually help: **Be explicit about format.** Vague prompts invite interpretation. If you want a five-sentence summary in plain language, say that. The more specific the instruction, the less room the model has to improvise. **Specify the role and constraints up front.** Telling the model to respond as a specific type of expert, to avoid certain topics, or to follow a particular structure gives it guardrails that reduce drift between runs. **Start fresh conversations for unrelated tasks.** Accumulated context is a source of variation. If you're running the same prompt repeatedly, a clean session removes the conversational baggage that can nudge the output in unexpected directions. **Save outputs you like rather than regenerating.** For anything where consistency matters, the most reliable approach is to keep the specific output you want and reuse it. Regenerating introduces a new probabilistic event every time. **Build a prompt library.** Saving and reusing effective prompt wording is one of the most practical ways to maintain consistency across sessions. If a particular phrasing reliably gets you what you need, treat it as an asset. --- ## The underlying issue is usually the prompt Most of the time, when someone gets wildly different answers to the same question, the prompt is doing less work than they think. A short, open-ended question gives the model enormous latitude. It can answer from multiple angles, at different levels of detail, in different tones, and all of those answers might be technically correct. The variation isn't a bug, it's the model filling in the gaps you left. Tightening the prompt, specifying the format, naming the constraints, and being explicit about what you want tends to narrow the range of outputs significantly. This is harder than it sounds, because most people write prompts the way they'd ask a question to a person who already knows their context. The model doesn't have that context unless you give it. This is exactly what [FixMyPrompt](https://fixmyprompt.net/try) is built for. Paste your prompt, get a score from 0 to 100 and see every weak spot flagged with a specific fix, free, no signup. A free account adds the rewritten version. If your outputs are inconsistent, the diagnosis usually starts with the prompt. --- ## Sources - WSU study on ChatGPT response consistency: https://www.wsu.edu (2026 study on consistency across 10 identical prompts) - OpenAI documentation on temperature and sampling parameters: https://platform.openai.com/docs/api-reference - OpenAI model updates and versioning: https://platform.openai.com/docs/models ## Sources - https://multiple.chat/why-ai-answers-inconsistent - https://scalemath.com/blog/does-chatgpt-give-the-same-answers-to-everyone - https://community.openai.com/t/by-giving-the-same-input-how-can-i-make-chat-gpts-output-more-stable-or-consistency/428228 - https://promptly.fyi/blog/why-ai-gives-different-answers - https://www.reddit.com/r/PromptEngineering/comments/1nyprx5/why_does_the_same_prompt_give_me_different - https://medium.com/@macplanet2012/why-chatgpt-gives-you-a-different-answer-every-time-its-not-randomness-00d86dbcfe13 --- ## How Do You Ask ChatGPT to Teach You a New Language? Published: 2026-09-07 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/how-to-ask-chatgpt-to-teach-you-a-language Learn a reusable daily lesson prompt for ChatGPT with level tracking and correction rules, and why one-shot prompts never work for language learning. You decided to learn Spanish. Or French. Or Japanese. You opened ChatGPT, typed "teach me Spanish," and got a wall of text you read once and never went back to. A week later, nothing stuck. That is not a you problem. It is a setup problem, and this post fixes it. If you have never used ChatGPT before, here is how to start: go to chatgpt.com and make a free account, it only takes a minute. Once you are in, you will see a box at the bottom of the screen. Type there and press Enter. That is the whole thing. --- ## Why "teach me Spanish" does not work When you type something short and vague, ChatGPT has almost nothing to go on. It does not know your level, how much time you have, or what kind of help you want. So it guesses, usually toward a generic overview that fits nobody in particular. This is what researchers who study these tools call a one-shot approach: you give one bare instruction, and the tool gives one generic response. It sometimes makes things up, it rarely builds on what you said before, and there is no structure to come back to tomorrow. Each morning you would have to start from scratch. The fix is to give it a proper set of instructions once, then reuse them every day. --- ## The seven things your daily lesson needs A well-built set of instructions for language learning has seven parts. Think of them as filling out a form before your first lesson with a tutor. **Role.** Tell ChatGPT to act as your language coach. This shapes the tone of everything it says back to you. **Level.** Tell it where you are starting. Complete beginner, a little rusty from high school, somewhere in the middle. Be honest. **Situation.** Give it a scene to practice. Ordering coffee. Asking for directions. Introducing yourself at a dinner party. Specific scenes produce useful practice. **Constraint.** Set limits on complexity. If you are a beginner, you do not want it throwing advanced grammar at you in the first week. **Feedback rule.** Decide how you want to be corrected. Some people want every mistake caught. Others find that overwhelming and learn better when only the most important error gets flagged each time. Either is fine, just decide. **Output format.** Tell it how to lay out the answer. A vocabulary table, example sentences, a short dialogue, a grammar table with positive, negative, and question forms side by side. Specific formats are easier to study from. **Reuse task.** Add a line that tells it to end every lesson with a prompt for tomorrow. This is what turns a one-time answer into a daily habit. --- ## Your copy-paste daily lesson block Here is a set of instructions you can use every morning. Copy the whole thing, then swap out the parts in [square brackets] for your own details. ``` You are my personal [Spanish] coach. I am a [complete beginner]. Today's situation: [ordering food at a café]. Keep vocabulary and grammar at beginner level. No more than [ten] new words per lesson. Correct my mistakes one at a time, starting with the most important one. Format the lesson as: a short vocabulary table, two example sentences I can say out loud, and one short dialogue I can practice. At the end, suggest tomorrow's situation so I can paste it back in. ``` Here is what that looks like filled in with real details: ``` You are my personal French coach. I am a complete beginner. Today's situation: buying bread at a bakery. Keep vocabulary and grammar at beginner level. No more than eight new words per lesson. Correct my mistakes one at a time, starting with the most important one. Format the lesson as: a short vocabulary table, two example sentences I can say out loud, and one short dialogue I can practice. At the end, suggest tomorrow's situation so I can paste it back in. ``` Paste that in, press Enter, and your lesson starts. You cannot break anything. If the answer comes back in a way that does not suit you, just tell it, "that was too complicated, can you simplify it?" and it will try again. --- ## Tracking your level as you go ChatGPT does not remember your previous conversations by default, so each morning it starts fresh. That sounds like a problem, but you can work around it easily. At the top of your daily block, add one line that tells it where you left off. Something like: "I have been studying for two weeks and I can handle present tense but not past tense yet." That single sentence brings it up to speed. ChatGPT also has a setting called Custom Instructions, where you can save your learning preferences once so they apply to every conversation automatically. You can write your level, your correction preference, and the kind of examples you like there, and you will not have to repeat yourself every day. Over thirty days of daily lessons, this kind of structure can take you from knowing nothing to handling basic conversations. That is not a guarantee, it depends on how much time you put in, but the structure makes consistent practice much easier than trying to remember where you left off with flashcards or a notebook. --- ## When you want to work on grammar specifically Some mornings you will want to focus on one grammar rule rather than a full lesson. For that, adjust the format line to ask for a table with positive, negative, and question forms, plus two or three example sentences for each. Seeing the same structure three ways in one place is much easier to absorb than reading a paragraph of explanation. You can also ask it to explain the rule in plain language first, then show the table. Tell it your level again in that message so it pitches the explanation right. --- ## A word about few-shot prompting You may come across advice to give ChatGPT several examples of what you want before asking your question. This is called few-shot prompting, and research has shown it helps with more complex tasks because the examples show the tool exactly what kind of answer you are after. For language learning, this means you could paste in one or two example lessons you liked and say "give me something like this." It takes a little more setup, but the results tend to be more consistent. --- ## Closing The difference between a lesson that helps and one that goes nowhere is almost always in how clearly you set it up. A few extra sentences at the start of your message, the same ones every morning, is all it takes to turn a vague chat into something that actually builds on itself. If you would like your instructions checked and scored automatically, fixmyprompt.net/try does it free, three times a day. --- ## Sources The structural framework and verified facts used in this post draw from the following sources provided for this article. No external statistics or third-party claims have been added beyond what was supplied. - Min, S., et al. (2022). Research on few-shot prompting and its benefits for complex tasks. *(Referenced in the verified facts supplied for this post; no direct URL provided.)* ## Sources - https://funfluen.com/learn/guides/chatgpt-prompts-for-language-learning - https://www.facebook.com/groups/aifire.co/posts/1618427175429252 - https://www.bgr.com/1997206/chatgpt-expert-tips-help-learn-new-language - https://learnprompting.org/blog/ai-prompt-language-learning - https://www.promptingguide.ai/techniques/fewshot - https://medium.com/@andriizolkin/chatgpt-zero-vs-few-vs-one-shot-prompting-eec8a1b7c055 --- ## Why Does ChatGPT Ignore My Instructions Even When It Understands Them? Published: 2026-09-03 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/chatgpt-ignores-my-instructions ChatGPT often understands your instructions but still breaks them. Learn why negative constraints leak, instruction burying happens, and how to fix both. You gave ChatGPT a clear instruction. It acknowledged the instruction, maybe even repeated it back to you, and then ignored it completely. This is one of the more frustrating experiences in working with these models, because it looks like a comprehension failure when it usually isn't. The model understood you. It just didn't comply. Here's why that happens, and what you can do about it. --- ## The Model Is Trained to Be Helpful, Not Obedient ChatGPT's training process uses Reinforcement Learning from Human Feedback (RLHF), which means human raters shaped the model's behavior by rewarding responses that felt helpful and agreeable. Over time, that reward signal baked in a strong bias: when the model faces a choice between following your instruction strictly and producing something that feels more complete or useful, it often drifts toward the latter. This is a design outcome, not a bug. The model was optimized to satisfy, and "satisfying" often means adding context you didn't ask for, softening a constraint you stated clearly, or reverting to a default behavior that raters historically rewarded. When your instruction conflicts with that trained instinct, the trained instinct frequently wins. The practical consequence: if you tell ChatGPT "do not explain the background, just give me the answer," it may give you the answer and then explain the background anyway. It understood the constraint. It overrode it because its training pushes toward thoroughness. --- ## Negative Constraints Leak Negative instructions, the "do NOT do X" variety, are harder for the model to hold than positive ones. A positive instruction gives the model a target to move toward. A negative instruction asks it to suppress a behavior that its training actively reinforces. This doesn't mean negative constraints are useless. Hard phrasing matters: "Do NOT include bullet points" performs better than "try to avoid bullet points." The softer version reads as a preference the model can weigh against other factors. The harder version reads as a rule, and the model treats it more like one. Even so, negative constraints degrade over a long conversation. As the chat grows, earlier instructions carry less weight because the model attends more heavily to recent tokens. A constraint you set in message one may be functionally invisible by message fifteen. This is why you'll sometimes give corrective feedback, the model acknowledges it, and then two responses later the same behavior reappears. The acknowledgment was real. The retention wasn't. --- ## Instruction Burying Where you place an instruction inside a prompt changes how much the model respects it. Instructions buried in the middle of a long prompt, surrounded by context, examples, and qualifications, get less attention than instructions placed at the start or the end. The fix is mechanical: put your most important constraints first, before any background or context, and repeat the critical ones at the end. This isn't about the model being lazy. It's about how attention works across a long sequence of tokens. The beginning and end of a prompt are structurally more prominent. Custom Instructions in ChatGPT compound this problem. Each of the two fields is capped at 1,500 characters, and as a conversation extends, those instructions lose influence to the more recent messages in the thread. If your system-level rules live only in Custom Instructions, they will erode. Task-specific prompts that conflict with your Custom Instructions will often win, because they appear later in the context window. --- ## When Instructions Conflict, the Model Defaults to Agreeable ChatGPT doesn't have a clean conflict-resolution mechanism that surfaces to the user. When it encounters two instructions that pull in different directions, it doesn't flag the conflict and ask you to resolve it. It defaults to its base training, which is to produce something agreeable and complete. This creates a specific failure pattern: you set a rule in Custom Instructions, then you write a task prompt that implicitly contradicts that rule, and the model follows the task prompt while quietly dropping the rule. You may not notice until the behavior you were trying to prevent shows up again. The response to this is to audit your prompts for implicit conflicts before you run them. If your Custom Instructions say "always respond in plain text, no formatting" and your task prompt says "give me a structured breakdown of these five points," you've created a conflict. The model will probably give you formatting. --- ## Long Conversations Break Down Faster Than You Expect Consistency degrades in long threads. The more constraints you've accumulated, the more likely they are to conflict with each other in ways you didn't anticipate, and the more likely the model is to start making its own judgment calls about which ones to honor. A few habits that help: - Start a new conversation when the task changes significantly, rather than continuing a thread that's carrying a lot of accumulated context. - Restate your most important constraints at the top of each major new prompt within a session, rather than assuming they're still active. - Keep your constraint list short. A prompt with twelve rules will see more rule-breaking than a prompt with three, because the model has more opportunities to trade one constraint against another. --- ## What Structured Prompts Actually Do "Write a structured prompt" is advice that gets repeated without much explanation of the mechanism. Here's what it actually changes: a structured prompt separates your constraints from your context from your task, so the model can parse each component clearly. When everything runs together in a paragraph, the model has to infer what's a rule and what's background. When you separate them explicitly, that inference step goes away. A minimal structure that works: **Constraints** (listed first, hard language) **Context** (what the model needs to know) **Task** (what you want it to produce) This isn't a magic format. It's a way of reducing the ambiguity that gives the model room to drift. --- ## The Underlying Issue Instruction-following in ChatGPT is not purely a prompt-writing problem. The model's conversational state tracking is limited, its training creates genuine tension between compliance and helpfulness, and long threads accumulate enough noise that even well-written instructions lose their grip. You can write better prompts and get meaningfully better results, but you're working with a model that was shaped to satisfy rather than to obey. Knowing that changes how you approach it. You're not trying to write a perfect instruction that the model will follow forever. You're managing attention, reducing ambiguity, and minimizing the gaps where the model's defaults can fill in. --- If bad output keeps following you across tools and models, the prompt is usually where the problem starts. FixMyPrompt scores your prompt from 0 to 100, flags the specific weak spots (buried constraints, conflicting instructions, soft language where hard language would work better), and rewrites it. [Try it at FixMyPrompt.net/try.](https://fixmyprompt.net/try) --- ## Sources 1. OpenAI. "Reinforcement Learning from Human Feedback (RLHF)." https://openai.com/research/learning-to-summarize-with-human-feedback 2. OpenAI Help Center. "Custom Instructions for ChatGPT." https://help.openai.com/en/articles/8096356-custom-instructions-for-chatgpt 3. Anthropic. "Core Views on AI Safety." https://www.anthropic.com/index/core-views-on-ai-safety 4. OpenAI. "GPT-4 Technical Report." https://arxiv.org/abs/2303.08774 5. Perez, E., et al. "Red Teaming Language Models with Language Models." https://arxiv.org/abs/2202.03286 ## Sources - https://resources.opencraftai.com/blog/why-chatgpt-keeps-ignoring-custom-instructions-and-what-actually-works - https://aitoolsusageguide.org/chatgpt-ignores-instructions - https://www.memorylake.ai/en/blogs/chatgpt-forgets-custom-instructions - https://docs.bswen.com/blog/2026-02-22-why-chatgpt-ignores-instructions - https://community.openai.com/t/catastrophic-failures-of-chatgpt-thats-creating-major-problems-for-users/1156230 - https://community.openai.com/t/assistant-ignores-latest-instructions-and-overrides-user-constraints-in-edits/1379048 --- ## Why Does ChatGPT Stop Mid-Response and How Do You Fix It? Published: 2026-08-31 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/why-does-chatgpt-stop-halfway ChatGPT cuts off for five distinct reasons. Learn how to diagnose which cause is affecting you and which fixes actually work for each one. You asked ChatGPT for a detailed explanation, a long piece of code, or a structured breakdown, and it stopped in the middle of a sentence. Maybe it cut off right before the conclusion. Maybe it stopped halfway through a function. You typed "continue" and it picked up, but the code it resumed with broke something that was working before. You tried again. Same result. This is one of the most common frustrations with ChatGPT, and the fix depends entirely on which of several distinct causes is actually responsible. Treating them as the same problem is why most workarounds fail. ## The five reasons ChatGPT stops mid-answer Cut-off responses come from five separate causes, and they look almost identical on the surface. **The output-token ceiling.** Every response has a maximum number of tokens it can generate before the model simply stops. A token is roughly three to four characters, so a long technical answer or a substantial block of code can hit this ceiling before the content is complete. This is the cause most people assume is always responsible, but it is only one of five. **A full context window.** The context window is the total amount of text, both your inputs and ChatGPT's outputs, that the model can hold in working memory for a given conversation. When that fills up, the model loses access to earlier parts of the thread. Responses can become truncated or incoherent. Crucially, if this is the cause and you respond with "keep going," you are adding another turn to an already full thread, which tends to make the situation worse rather than better. **A dropped connection.** ChatGPT streams its responses over a websocket. If that connection drops mid-generation, the response stops wherever it was when the connection broke. This looks exactly like hitting the token ceiling, but the underlying cause is a network or server issue, and the fix is different. Refreshing and regenerating often resolves it; restructuring your prompt does not. **Rendering failures.** Sometimes the full response was generated but the interface failed to display all of it. This is a limitation of the chat interface rather than the model itself, and it can occur with very long outputs or complex formatting. **A prompt that implies brevity.** This one surprises people. If your prompt is phrased in a way that signals a short answer is expected, the model may produce one. Vague or casual phrasing can lead to a response calibrated to match the apparent scope of the question, not the scope you actually needed. ## Why "continue" often makes things worse Typing "keep going" or "continue where you left off" works sometimes, specifically when the cause was the output-token ceiling and the context window still has room. In that situation, the model can pick up from roughly where it stopped. The problems start when the cause was something else. If the context window is full, continuing the conversation adds more text to an already overloaded thread. The model may repeat itself, skip sections, or produce a response that contradicts what it said earlier. For code specifically, asking ChatGPT to resume a truncated function carries real risk. The model may rewrite sections it already completed, alter variable names, or change logic in ways that break functionality that was previously correct. If you are working with a large code block, this is worth treating as a distinct hazard rather than a minor inconvenience. Typing a single space and pressing enter is a workaround some users report as effective. The "Regenerate response" button can also prompt ChatGPT to retype a response that stopped unexpectedly. These are useful when the cause was a connection drop or a rendering failure, but they do not address structural causes. ## How to diagnose which cause you are dealing with Before trying a fix, spend a moment identifying what actually happened. Did the response stop at a consistent length across multiple attempts? That points to the token ceiling or a prompt-level issue. Did it stop at an unpredictable point, or did the interface behave strangely? That suggests a connection or rendering problem. Is this a long conversation where earlier context seems to be getting lost? The context window is likely filling up. Server outages and glitches are also a real cause. If ChatGPT is cutting off responses across different prompts and different conversation lengths, checking the OpenAI status page before spending time debugging your prompt is worth doing. ## Prompt structures that prevent cut-offs When the cause is the token ceiling or a prompt implying brevity, the structure of your request matters considerably. Asking for everything at once in a single prompt is the most common setup for a cut-off. A prompt like "write a complete guide to X" gives the model no natural stopping points and no signal about how to pace the output. When the ceiling arrives, the response just ends. A few structural changes reduce this reliably. Ask for one section at a time. Instead of requesting the full document, ask for the introduction, confirm it, then ask for the next section. This keeps each individual response well within the output ceiling and gives you checkpoints to catch problems early. Be explicit about length and format. A prompt that specifies "write approximately 400 words on X, covering these four points" gives the model a calibrated target. Vague prompts produce responses calibrated to the model's interpretation of your intent, which may be much shorter than what you needed. For code, ask for the function in pieces if it is long. Request the data model first, then the logic, then the error handling. Asking for a complete working application in one prompt is the fastest route to a truncated, partially broken output. Name the deliverable clearly. "Explain" implies a shorter response than "write a step-by-step guide." "Summarize" implies something different from "walk me through." The words you use to frame the request shape what the model treats as a complete answer. ## The underlying issue most cut-offs share A dropped connection is outside your control. A server glitch is outside your control. But the majority of cut-offs that happen repeatedly, across different sessions, on similar types of requests, trace back to a prompt that did not give the model enough structure to produce a complete response. Vague prompts produce responses that are either too short, cut off at an arbitrary point, or resumed in ways that introduce errors. The model is not being uncooperative; it is working with the information you gave it. If you are getting cut-offs regularly and the workarounds are not holding, the prompt itself is usually where the problem lives. FixMyPrompt scores your prompt, flags the structural weaknesses that lead to truncated or generic responses, and rewrites it. You can [try it at fixmyprompt.net/try](https://fixmyprompt.net/try). ## Sources - https://www.reddit.com/r/ChatGPT/comments/11sxx3w/why_does_chatgpt_stop_generating_text/ - https://www.makeuseof.com/why-does-chatgpt-stop-generating-text/ - https://community.openai.com/t/why-does-chatgpt-cut-off-mid-response/ - https://prompt-architects.com/blog/140-why-does-chatgpt-cut-off-mid-answer - https://www.wikihow.com/Why-Does-Chat-Gpt-Stop - https://duttonlaw.ca/chat-gpt-stop-outputting-mid-sentence-solved - https://www.reddit.com/r/GPT3/comments/zgmank/chatgpt_responses_getting_cut_off - https://community.openai.com/t/chatgpt-sometimes-not-finish-its-code-or-sentence-how-do-i-solve-this/46476 - https://community.pickaxeproject.com/t/chat-gpt-4o-responses-are-getting-cut-off-fairly-quickly/417 --- ## How Do You Ask ChatGPT to Create a Cookbook? Published: 2026-08-31 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/how-to-ask-chatgpt-to-create-a-cookbook Learn how to ask ChatGPT to create a family cookbook step by step. Get a chapter outline, write recipes with family stories, and assemble your book using free tools. You have a box of recipe cards, a head full of family stories, and a real wish to turn them into something your grandchildren can actually use. You sit down, open ChatGPT, and type "make me a family cookbook." It gives you a list of recipes. You wanted a book. That gap is exactly what this post closes. --- ## First, here is how to open ChatGPT if you have never done it Go to chatgpt.com and make a free account. It takes about two minutes and only needs an email address. Once you are in, you will see a box at the bottom of the screen. Type there, press Enter, and ChatGPT replies. That is the whole thing. You cannot break anything, and if the answer is not what you wanted, you can just tell it that and ask again. --- ## Why a vague request gives you a list instead of a book ChatGPT is very good at writing individual recipes when you ask for them. What it does not have is a built-in system for organizing a whole cookbook. It does not automatically know how many chapters you want, what order they should go in, or how each recipe page should look. So if you ask in a general way, it does the general thing: it lists recipes. The fix is to describe your cookbook before you ask for a single recipe. Think of it like calling a carpenter. "Build me something wooden" gets you a birdhouse. "Build me a six-drawer dresser, oak, 48 inches tall, with dovetail joints" gets you a dresser. --- ## Step one: describe your cookbook in one clear block Here is the most important thing you will do. Write out your cookbook idea in detail before you ask for anything else. The more specific you are, the more useful the answer will be. Copy the block below, then change the parts in [square brackets] to match your own cookbook: ``` I want to create a family cookbook. Here are the details: Cuisine type: [e.g. Southern American comfort food] Who it is for: [e.g. my grandchildren, ages 8 to 16, cooking with a parent] Number of recipes: [e.g. 30 recipes] Dietary focus: [e.g. no nuts, low sugar] Themes or sections: [e.g. Sunday breakfasts, weeknight dinners, holiday desserts] Each recipe page should include: [e.g. a short family story, ingredient list, step-by-step instructions, and a tip for beginners] Tone: [e.g. warm and simple, like a grandmother explaining it out loud] Please start by giving me a full chapter outline for this cookbook, with section names and a short description of each section. ``` Here is what that looks like filled in with real details: ``` I want to create a family cookbook. Here are the details: Cuisine type: Southern American comfort food Who it is for: my grandchildren, ages 8 to 16, cooking with a parent Number of recipes: 30 recipes Dietary focus: no nuts, low sugar Themes or sections: Sunday breakfasts, weeknight dinners, holiday desserts Each recipe page should include: a short family story, ingredient list, step-by-step instructions, and a tip for beginners Tone: warm and simple, like a grandmother explaining it out loud Please start by giving me a full chapter outline for this cookbook, with section names and a short description of each section. ``` Once you have your outline, save it somewhere, a Word document, a notes app, anywhere. You will refer back to it constantly. --- ## Step two: work through one section at a time Do not ask for all 30 recipes at once. Ask for one section at a time, or even one recipe at a time. Smaller requests give you better, more careful answers. When you start a new session on a different day, paste your outline back in at the top so ChatGPT remembers what you are building. It can lose track of earlier conversations, so giving it that reminder keeps everything consistent. For each recipe, you can say something like: "Using the cookbook outline above, write the Sunday Breakfasts chapter. Start with the first recipe: buttermilk biscuits. Include a short family story, the ingredient list, step-by-step instructions, and a tip for beginners. Use a warm, simple tone." If the answer comes back too complicated, just say "can you make the instructions simpler?" It will rewrite them. If you want more detail, say "can you add a note about what to do if the dough is too sticky?" It adjusts. --- ## Step three: assemble the pages yourself ChatGPT will not package everything into a finished file for you. It does not export to PDF or create a designed book. What it gives you is text, and you copy that text into wherever you are building your book. A free and friendly option is Canva (canva.com). It has cookbook templates you can find by searching "cookbook" in the template library. You paste your recipe text into the template, add any photos you have, and adjust the look to your taste. Canva's Bulk Create feature can even help you build many pages at once if your recipes follow a consistent format, which they will if you asked ChatGPT to use the same structure every time. Yes, this is manual work. You are building page by page. But you are also making something genuinely yours, with your family's stories and your own choices about how it looks. --- ## A note if you want more automation If the copy-paste process feels like too much, there are tools built specifically for cookbook creation that handle the structure and export for you. One is Inkfluence AI, which offers structured outlines, consistent recipe formatting throughout, and PDF and EPUB export as part of its paid plans. That is a different kind of tool for a different kind of need. For a one-time family project, the ChatGPT and Canva path works well. For someone making cookbooks regularly, a dedicated tool may save time. --- ## The short version of the whole process Describe your cookbook fully before asking for anything. Get your chapter outline first. Then work through one section at a time, pasting your outline back in whenever you start a new session. Copy each finished recipe into a Canva template. Repeat until the book is done. The describing-first step is the one most people skip, and it is the reason they end up with a list instead of a book. --- If you would like your instructions checked and improved automatically before you start, fixmyprompt.net/try does it free, three times a day. --- ## Sources - https://www.inkfluenceai.com - https://www.canva.com/features/bulk-create/ - https://help.openai.com/en/articles/6783457-what-is-chatgpt ## Sources - https://chriskresser.com/turn-chatgpt-into-your-personal-meal-planner - https://www.tiktok.com/discover/how-to-write-a-cook-book-using-chatgpt - https://estherjacobs.info/en/blog/how-i-wrote-a-book-in-one-week-with-chatgpt - https://www.inkfluenceai.com/write/cookbook - https://medium.com/@danalettix/how-i-format-my-cookbook-recipe-pages-using-chatgpt-and-canva-and-produce-dozens-at-a-time-7c2f81f29010 - https://www.reddit.com/r/ChatGPTPro/comments/114in3j/i_wrote_two_100_page_books_using_chatgpt_a_recipe --- ## Why Does ChatGPT Give Such Generic Answers? (And the Fix) Published: 2026-08-29 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/chatgpt-generic-answers ChatGPT gives vague, generic answers when the prompt leaves gaps it fills with averages. Here is why it happens, the five gaps that cause it, and how to get specific, detailed answers instead. You ask ChatGPT something you actually care about, and it hands back the kind of answer that could have been written for anyone. Five bullet points you already knew. Advice so general it fits every situation and helps in none. "Consider your audience." "Consistency is key." You did not need an AI for that, and the frustrating part is watching other people apparently get sharp, specific, useful answers from the same tool. Here is the thing nobody tells you when you start: the model is not being lazy, and yours is not broken. A generic answer is what a language model produces when the question leaves it nothing to be specific *with*. The gaps in your prompt get filled with statistical averages, and the average of everything is beige. That is genuinely good news, because it means the fix is on your side of the screen and takes about a minute. ## What "generic" actually means to a language model ChatGPT predicts the most likely helpful response to what you wrote. Most likely is the operative phrase. When you ask "how do I market my business," the model has no idea whether you run a nail salon in Ohio or a B2B software company in Berlin, so it produces the response most likely to be acceptable across all of them: post on social media, know your audience, be consistent. That answer is not wrong. It is the mathematical center of every marketing answer ever written, which is exactly why it feels like it came from a fortune cookie. The people getting impressive answers are not using a different ChatGPT. They are asking questions that pin the model to their specific situation, so the "most likely response" becomes the most likely response *for that situation* instead of for the average of all situations. ## The five gaps that produce a generic answer Nearly every vague answer traces to one of these missing pieces. Read your last disappointing prompt against this list and you will usually find two or three of them open. **1. Who is asking.** "How do I negotiate a raise" gets the average of all raise advice. "I'm a nurse with six years at the same hospital, top performance reviews, and an offer from a competitor. How do I negotiate a raise" gets a strategy. The model calibrates everything, from vocabulary to which advice even applies, based on who it thinks it is talking to, and when you do not say, it assumes nobody in particular. **2. What the answer is for.** The same question needs a different answer depending on whether you are writing an email, making a decision this afternoon, or learning a topic from scratch. "Explain inflation" and "explain inflation so I can decide whether to move savings into I-bonds this month" are different requests wearing the same words. **3. What you already know.** Without this, the model plays it safe and starts from zero, which reads as condescending filler if you are past zero. One sentence fixes it: "I already understand X, skip the basics." **4. What a good answer looks like.** Length, format, depth, examples or no examples. The model has a default for each, and its defaults are tuned for the average person, not for you. If you want three concrete options with tradeoffs instead of eight shallow bullets, that sentence has to exist in the prompt. **5. The constraints that make it real.** Budget, deadline, tools you already use, things you have already tried and ruled out. Constraints are what turn advice from theoretical to usable, and the model cannot guess a single one of them. ## The before and after, on a real prompt Before: > Give me ideas for content for my business. That prompt could belong to anyone on earth, so the answer will too. After: > I run a two-person bookkeeping firm serving restaurant owners in Texas. I post on LinkedIn twice a week; my goal is getting restaurant owners to book a free consult. My last three posts about tax deadlines got good engagement but no bookings. Give me 5 post ideas that pull toward a consult, and for each one say why a restaurant owner would care. Don't suggest video, I won't make videos. Same tool, same day, same model. The second prompt gets a response that reads like it came from a marketing consultant who knows the business, because every sentence in it closed a gap the model would otherwise fill with an average. Notice what the fix was not: no magic phrases, no "act as an expert" incantation, no prompt-engineering jargon. Just the information a competent human would also have needed before giving you specific advice. ## Why "be more detailed" doesn't fix it The instinctive repair is to add "be specific" or "give me details" to the prompt, and it mostly fails. The model cannot be specific about facts it does not have. Told to be detailed with nothing to be detailed *about*, it produces the same generic answer with more words in it, which is arguably worse. Specificity in equals specificity out; there is no instruction that substitutes for the missing information. The other instinctive repair is blaming the model and switching to Claude or Gemini. Sometimes model choice genuinely matters, and we have written about [where Gemini goes wrong](/blogs/15-gemini-wrong-answers) and how the models differ. But a vague prompt is vague on every model. Switching tools to escape a vague prompt just hands the same gaps to a different gap-filler. ## The 60-second pass that fixes most prompts Before sending a prompt that matters, add one sentence for each of these four, in plain language: 1. **Who you are and what situation you are in.** 2. **What the answer is for.** What you will do with it. 3. **What good looks like.** Length, format, depth. 4. **What is off the table.** Constraints, things tried, things you refuse to do. That is not prompt engineering. It is the same context you would give a human expert in the first two minutes of a conversation, and the model needs it for the same reason. If the answer still comes back generic after this, the problem is usually deeper in the prompt structure, and this related pattern is worth checking: ChatGPT [ignoring or misreading instructions](/blogs/07-chatgpt-misunderstanding-instructions) has different causes than vagueness and a different set of fixes. And if answers seem to have gotten blander across the board recently rather than for one prompt, that is a real phenomenon with its own explanation: see [what actually changed in ChatGPT this year](/blogs/19-is-chatgpt-getting-worse-2026). ## When you'd rather not do this by hand Everything above works, and plenty of people read this far and still will not do it, because rereading your own prompt for five kinds of missing context is exactly the sort of chore humans skip when the deadline is now. That is the case [FixMyPrompt](/try) was built for. Paste the prompt, and it scores it 0 to 100 and flags each gap of the five above by name. The generic-answer problem is visible in the score before you waste the run: prompts that produce beige answers reliably score in the 40s and 50s, and the per-criterion breakdown shows exactly which of the gaps above is doing the damage. Three free checks a day, no signup. A free account unlocks the rewrite, which hands back a version with the gaps closed for you to fill in with your specifics. ## The short version ChatGPT gives generic answers when the prompt leaves gaps, because a language model fills every gap with the statistical average, and the average of everything is vague. The five gaps that matter: who is asking, what the answer is for, what you already know, what a good answer looks like, and the real-world constraints. Close them in plain sentences and the same model produces specific, usable answers. "Be more specific" does not work, because there is no instruction that substitutes for missing information. And if you would rather not audit your own prompts, [a prompt checker](/try) does the audit and the rewrite in one pass. --- ## Claude Sonnet 5 Pricing and What Changed from 4.6 Published: 2026-07-05 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/claude-sonnet-5-guide Claude Sonnet 5 is Anthropic's new speed-and-intelligence model, at $3/$15 per million tokens with an introductory rate of $2/$10 through August 2026. Here is what changed from Sonnet 4.6, the exact pricing and limits, and how to run your own prompt against it before you commit. Anthropic released Claude Sonnet 5 (`claude-sonnet-5`) as the new middle tier of the Claude lineup, the one it describes as the best combination of speed and intelligence. If you were using Sonnet 4.6 for everyday work, this is the model that replaces it at the top of your list. Here is what actually changed, what it costs, and how to see its output on your own prompt before you switch anything. ## The pricing, plainly Sonnet 5 lists at **$3 per million input tokens and $15 per million output tokens**. That is the same headline rate as Sonnet 4.6, so a straight switch does not change your per-token math. The part worth knowing is the introductory rate. Through **August 31, 2026**, Sonnet 5 runs at **$2 input and $10 per million output**, roughly a third cheaper than the standard rate. After that date it returns to $3/$15. If you have volume to move, the window before September is the cheap time to do it. ## What changed from Sonnet 4.6 The context window and output cap are unchanged: a **1 million token context** and up to **128,000 output tokens** on the standard Messages API. So anything that fit in Sonnet 4.6 fits here. Two differences matter in practice: - **Adaptive thinking is on, and effort defaults to high.** On the Claude API and in Claude Code, Sonnet 5 sets its effort level to high unless you tell it otherwise. That buys better answers on harder prompts, and it can also mean slower, more expensive responses on simple ones than you saw by default on older Sonnets. If you run high volume of easy calls, set the effort down. - **A newer knowledge cutoff.** Sonnet 5 has reliable knowledge through January 2026, later than Sonnet 4.6. For anything time-sensitive that still is not enough, and you should give the model the current facts in the prompt rather than trusting recall. Anthropic has not published head-to-head numbers that we would repeat as fact, so the honest move is to judge it on your own work rather than a benchmark chart. Which brings us to the useful part. ## How to test Sonnet 5 on your own prompt in FixMyPrompt A model comparison only means something on the prompt you actually run. FixMyPrompt lets you take your prompt and run it against Sonnet 5 directly, so you see the real output instead of a leaderboard. 1. Open your prompt in FixMyPrompt. 2. Pick **Claude Sonnet 5** from the model list in the test run. 3. Run it, and read the output for your actual task. Run it against Sonnet 4.6 or another model in the same place if you want a side by side. It is pay as you go, so a test run costs a small number of credits rather than a subscription. You get the answer to the only question that matters, which is whether Sonnet 5 writes a better response to your prompt than what you use now, before you rewire anything downstream. ## Should you switch? If Sonnet 4.6 was your default, Sonnet 5 is the natural upgrade at the same standard price, and cheaper until September. The one thing to watch is the high effort default, which can quietly raise cost and latency on lightweight calls. Set effort deliberately, test your real prompts, and let the output decide. [Run your prompt against Claude Sonnet 5 in FixMyPrompt](https://fixmyprompt.net/try). No signup needed to try it. --- ## Claude Fable 5 Pricing vs Opus 4.8: Full Breakdown Published: 2026-07-05 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/claude-fable-5-guide Claude Fable 5 is Anthropic's most capable widely released model, at $10/$50 per million tokens with a 1M context window and always-on thinking. Here is the pricing, how it stacks against Opus 4.8, when the extra cost is worth it, and how to run your own prompt against it. Claude Fable 5 (`claude-fable-5`) is the top of Anthropic's lineup, the most capable model it has made widely available. It is built for long-running agents and the hardest reasoning work, and it is priced accordingly. Here is what it costs, how it differs from Opus 4.8, and the cases where paying for it actually pays off. ## The pricing, plainly Fable 5 costs **$10 per million input tokens and $50 per million output tokens**. That is double Opus 4.8's $5/$25, and more than three times Sonnet 5's standard rate. There is no introductory discount, so the sticker price is the price. It has a **1 million token context window** and up to **128,000 output tokens** on the standard Messages API, the same envelope as Opus 4.8. The size is not what you are paying extra for. The capability on hard tasks is. ## Fable 5 vs Opus 4.8 They share the context window and the output cap, so the choice is about intelligence per dollar, not room to work. - **Opus 4.8** is Anthropic's model for complex agentic coding and enterprise work, at $5/$25. For most demanding tasks it is the sensible default. - **Fable 5** sits above it, aimed at the longest-horizon agentic runs and the most difficult reasoning. Anthropic positions it as next-generation intelligence for long-running agents, at the cost of higher price and slower responses. Two behaviors are worth knowing before you send it a prompt. **Thinking is always on**, so Fable 5 reasons on every request and you cannot turn that off. And its safety classifiers can decline a request and hand it to a fallback model, which surprises people in week one. We wrote up exactly what triggers that and how to prompt around it in [Claude Fable 5: what it can do, why it refuses, and the prompt fixes that stop the Opus fallback](https://fixmyprompt.net/blogs/claude-fable-5-refusals-prompting). ## When the extra cost is worth it At five times Sonnet 5's price, Fable 5 is not the model to reach for by reflex. It earns its rate on a narrow set of work: long autonomous agent runs where a single mistake compounds across many steps, and reasoning problems at the edge of what other models can do. For a listing description, a cold email, or a summary, a cheaper model gives you the same result for a fraction of the credits. The trap is paying frontier rates for a task that never needed them. The way to avoid it is to compare the output on your own prompt rather than assume the priciest model is the right one. ## How to run your prompt against Fable 5 in FixMyPrompt FixMyPrompt lets you take your prompt and run it against Fable 5 directly, so you can see whether the top-tier model actually produces a better answer for your task, or whether Opus 4.8 or Sonnet 5 gets you there for less. 1. Open your prompt in FixMyPrompt. 2. Pick **Claude Fable 5** from the model list in the test run. 3. Run it, then run the same prompt against a cheaper model and compare the two outputs. It is pay as you go, so you spend a few credits to answer the question instead of committing to the most expensive model on faith. On the hard, high-stakes prompts where Fable 5 pulls ahead, you will see it. On the everyday ones where it does not, you will save the difference. [Run your prompt against Claude Fable 5 in FixMyPrompt](https://fixmyprompt.net/try). No signup needed to try it. --- ## The ChatGPT Prompt That Writes Etsy Listings That Actually Sell Published: 2026-06-29 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/23-etsy-listing-prompt Why are your Etsy listings not selling? Usually the title and tags don't match how buyers search. This is the prompt that fixes that, with a real before and after. You typed "write an Etsy description for my printable budget planner." ChatGPT gave you this: > Introducing our beautiful Budget Planner, the perfect tool to help you take control of your finances! This stunning printable is designed to make budgeting easy and enjoyable. Whether you are a beginner or a pro, our planner has everything you need to reach your financial goals. Download today and start your journey to financial freedom! It reads fine. It also will not sell, and the reason is not the writing. It is that nothing in there matches how an Etsy buyer actually searches, and Etsy's search is the only thing that decides whether your listing gets seen at all. A buyer does not type "beautiful budget planner stunning printable." They type "monthly budget planner printable" or "paycheck budgeting template digital download" or "dave ramsey style budget sheet." If those exact phrases are not in your title and tags, your listing does not show up, and a description this generic never gets a chance to convert anyone. ## Why the default version sinks The title is keyword-blind. ChatGPT wrote a description and never touched the title, which is the single biggest ranking factor on Etsy. Whatever you slapped on top is probably your product name, not a phrase anyone searches. The tags are missing. Etsy gives you 13 tags and weights them heavily. The default output ignores them completely, so you are leaving most of your search visibility on the table. The opening is wasted. The first line of your description shows up in Google results and is the first thing a buyer reads. "Introducing our beautiful..." tells them nothing about whether this planner fits what they need. There is no buyer in the room. The copy describes the product. It never names the person buying it, the moment they are buying it for, or the specific result they want. That is what turns a view into a sale. ## What a listing that ranks and sells looks like It starts from the words buyers type, not the words you would use to describe your own work. The clearest way to see it is the same product listed two ways. Here is a printable wall art listing that stays invisible: > Title: Boho Wall Art Print > > Tags: boho, wall art, print, home decor, cute, trendy, art, decor, wall, printable, digital, download, aesthetic > > First line: Add a touch of boho charm to your space with this gorgeous printable wall art! Nothing there matches a real search. "Boho wall art print" is what the seller calls it, not what a buyer types. The tags are single words that compete with a million other listings, and half of them just repeat the title. The first line talks about the product instead of the wall it is going on. Here is the same print built to be found: > Title: Boho Neutral Wall Art Printable, Beige Abstract Arch Print, Above Bed Decor, Digital Download > > Tags: neutral wall art, beige abstract print, above bed art, boho bedroom decor, arch print printable, minimalist living room, earthy tone wall art, digital download print, framed wall art set, ... > > First line: If you want to fill that blank wall above your bed without waiting on shipping, this is a printable set you can download and have framed tonight. Every phrase in the title is something a buyer actually types. The tags are multi-word and spread across styles (neutral, beige, minimalist), rooms (bedroom, above bed, living room), and format (printable, digital download), with no wasted repeats. The first line names the buyer's exact situation, a blank wall above the bed, and the result, framed tonight. The pattern holds for every product type. A wedding template, wrong then right: > Wrong: Wedding Invitation Template > > Right: Editable Wedding Invitation Template, Minimalist Greenery Invite, Canva Download, Printable Wedding Set The wrong title waits to be found for the one generic phrase "wedding invitation template," where it sits behind everyone. The right one stacks the specific phrases a bride searches: editable, minimalist, greenery, Canva, printable set. Same product, and only one of them shows up when someone is ready to buy. So a listing that ranks and sells front-loads the highest-intent phrase in the title, fills all 13 tags with real multi-word searches across styles and rooms and formats, and opens the description with the buyer's situation and the result rather than a description of the item. ChatGPT can do every bit of that. It just has to be told to, and told in the right order. Here is the prompt. ## The prompt > You are an Etsy SEO copywriter. I sell digital products and I need a listing that ranks in Etsy search and converts browsers into buyers. > > My product: [describe it in one line, e.g. "a printable monthly budget planner, digital download, minimalist black and white design"]. > My ideal buyer: [who they are and the moment they are shopping, e.g. "someone trying to get their spending under control this month, wants something simple they can print today"]. > > Before you write anything, list 10 to 15 exact phrases a real buyer would type into Etsy search to find this, ordered from highest buying intent to lowest. Use natural buyer language, not my product name. > > Then produce: > 1. A title, max 140 characters, that front-loads the single highest-intent phrase and works in 2 or 3 more naturally. No filler adjectives like "beautiful" or "stunning." > 2. Exactly 13 tags, each a multi-word phrase from real buyer searches. No single words, no duplicates, no repeating words already carried by the title. Cover variations, styles, occasions, and use cases. > 3. A description. The first two sentences name the buyer's situation and the result they get, and contain the top search phrase naturally. Then a short, scannable section on what is included and how it works. Plain, specific, no hype. > > After the listing, tell me the one phrase you would test changing first if it does not get views in two weeks. Two things make this work. It forces ChatGPT to start from buyer search terms before it writes a word, which is the step the default skips. And it gives the model the structure Etsy actually rewards, so the output is a listing you can paste in, not a paragraph you still have to rebuild. ## The before and after Same planner. Here is the difference in the part that decides everything, the title and the first line. Before: > Title: Budget Planner Printable > First line: Introducing our beautiful Budget Planner, the perfect tool to help you take control of your finances! After: > Title: Monthly Budget Planner Printable, Paycheck Budgeting Template, Digital Download Finance Tracker > First line: If you want your spending under control before this month gets away from you, this is a simple monthly budget you can print today and start using in ten minutes. The first one waits to be found and never is. The second one is built from the phrases buyers actually type, and the opening speaks to the person reading it instead of the seller who made it. ## The honest catch That prompt is long, and you do not want to retype it and re-explain your buyer every single time you list a new product. The whole reason you went to ChatGPT was to move faster, and a 200-word prompt you tweak for an hour is the opposite of fast. That is the gap FixMyPrompt closes. You give it your rough version, "write an Etsy description for my budget planner," and it hands back the upgraded, structured prompt above, tuned to your product, so you get the listing without building the prompt. It is not a content generator that competes with the free ones. It makes the prompt itself good, so whatever AI you already use produces a listing that gets found and sells. Your listing, in your words, sharpened to how buyers search. Try it on your worst-performing listing first. [Paste your rough prompt into FixMyPrompt](https://fixmyprompt.net/try), no signup, and see the difference in the title and tags before you change anything in your shop. --- ## Claude Fable 5 Is Back: What Happened and What Changed Published: 2026-06-22 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/claude-fable-5-unavailable-why-suspended Fable 5 launched June 9, was switched off June 12 by a US export-control order, and returned July 1 with a new safety classifier. Mythos 5 did not come back the same way. Here is the full timeline, what changed on the way back, and why your Fable prompts need re-checking. **Fable 5 is available again.** The US Department of Commerce lifted its export-control order on June 30, 2026, and Anthropic redeployed the model on July 1 to users worldwide. If you are reading this because you still see "Claude Fable 5 is currently unavailable," that message is stale on your end, so sign out and back in or check that your plan includes it. Mythos 5 is a different story, and the difference is the part most coverage skipped. It did not return alongside Fable. More on that below. I wrote the original version of this post while both models were dark and nobody had a return date. I have kept the history intact, because how this happened is still the most interesting part and it explains what changed on the way back. What follows is the timeline, why an order aimed at foreign nationals disabled the model for everyone, the jailbreak that set it off, what Anthropic changed before switching it back on, and what to do with the prompts you re-tuned during the outage. ## What happened, in order Fable 5 and Mythos 5 launched on June 9, 2026. They were the most capable models Anthropic had ever released to the public. Three days later they were gone. According to Anthropic's statement, the directive arrived on June 12 at 5:21pm ET. The reporting attributes it to Commerce Secretary Howard Lutnick. By that evening both models were disabled globally. The order was an export-control measure: it required Anthropic to "suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States." So this was a government action, not a product decision. That distinction matters, because it changes everything about how and when the model can come back. ## Why an order about foreign nationals took the model down for everyone The directive technically targeted foreign nationals. The shutdown hit every user. That gap confused a lot of people, so it is worth explaining the mechanism. Anthropic does not verify the nationality of each person sending a request in real time. There is no live filter that can let a US citizen through while blocking a foreign national on the next request to the same model. Faced with an order it had to comply with immediately and no way to enforce it selectively, the company took the only option that guaranteed compliance. In its own words, "the net effect of this order is that we must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance." That is why a US citizen in Ohio and a developer in Berlin both see the same "currently unavailable" message. The block is all-or-nothing because the enforcement tool is all-or-nothing. ## The jailbreak that triggered it The government's stated concern was national security, and the specific trigger was a demonstrated way to get around Fable 5's safeguards. Anthropic reviewed a demonstration of what it called "a narrow, non-universal jailbreak." The method involved asking the model to read a specific codebase and fix any software flaws in it. That sounds harmless, and in most contexts it is exactly the kind of work people want a frontier model to do. The security worry is the flip side of the same skill. A model good enough to find and fix flaws in a codebase is also a model good enough to find and exploit them, which moves it close to offensive cybersecurity capability. Reporting around the launch noted that Fable 5 had caught real bugs that GPT-5.5 and Opus 4.8 missed, so the capability was not theoretical. Anthropic was careful to frame the jailbreak as narrow rather than universal, meaning it affected specific tasks rather than unlocking the model wholesale. ## Where Mythos 5 fits in If you are new to the naming, Mythos 5 is the same underlying model as Fable 5 with the safeguards lifted, released to a small set of vetted partners. Fable 5 is the public version with safety classifiers watching the conversation. Because they share the same core capability, the directive named both, and both went dark together. Anyone who had partner access to Mythos 5 lost it at the same moment everyone else lost Fable 5. ## How Fable 5 came back The hold lasted eighteen days. Commerce lifted the export controls on June 30, Anthropic announced it was restoring access, and Fable 5 was live again on July 1 across the Claude Platform, claude.ai, Claude Code, and Cowork, for Pro, Max, Team, and select Enterprise plans. Access on Amazon Web Services, Google Cloud, and Microsoft Foundry followed. Usage was metered on the way back rather than thrown wide open. For the first week, through July 7, eligible subscribers could spend up to half of their plan's weekly usage limit on Fable 5. After that it moved to usage credits, which is where it sits now. The prediction in the original version of this post was half right. It was a regulatory hold that lifted when the regulator was satisfied, and Anthropic did close the jailbreak before relaunching. What I got wrong was the shape of the return. I expected a staged reopening starting with verified US enterprise accounts, and instead it came back globally on day one. ## What changed before it was switched back on Anthropic did not simply flip the same model on again. It trained a new safety classifier aimed at the bypass technique that caused the problem, and reports it catches those attempts in more than 99% of cases. The company also worked with the US government and industry partners on a shared jailbreak framework rather than treating the fix as its own patch. That number is Anthropic's own, measured against attacks they knew to test for, so read it as a strong claim rather than a settled fact. The practical consequence is what matters to you: **Fable 5 today refuses differently than Fable 5 did on June 9.** A new classifier watching a conversation changes where the line sits, and prompts that cleared the old one can trip the new one. If you built anything against launch-week Fable behaviour, it is worth re-checking rather than assuming it still runs. ## Mythos 5 did not come back the same way This is the question I still get asked most, and the answer is more specific than most coverage suggested. The Commerce order was lifted for both models, so headlines reasonably read "Fable and Mythos restored." Only one of them returned to the public. Anthropic enabled Mythos 5 for a set of US organizations on June 26, and access since then expands through **Glasswing**, its cybersecurity program that gives selected organizations frontier-model access for defensive security testing. So Mythos 5 is not off, and it is also not something you can go and use. It is gated behind a vetted program with a specific purpose, and the company has said it continues working with the US government to widen access to more domestic and international partners. If you are not in an organization doing defensive security work that qualifies for Glasswing, the practical answer to "when do I get Mythos 5" is that this is not the path it is on. Fable 5 is the version available to you, and for ordinary work it is the one you want anyway, since the difference is safeguards rather than raw capability. ## What to do with your prompts now Most people moved their prompts twice. Once onto Opus 4.8 or Sonnet 4.6 when Fable went dark, and again back toward Fable after July 1. Each of those moves quietly changed what good prompting looks like, and the second move is the one people skipped, because coming back to a model you used before feels like it should need no work. It does. A prompt reshaped for Opus 4.8 during the outage carries Opus-specific scaffolding that Fable 5 does not need, and Fable can read the extra structure as noise. Running it now costs you output quality and tokens at the same time, and it shows up as answers that are slightly flatter than you remember without any obvious cause. The new safety classifier is the second reason to re-check. Anything you wrote against launch-week Fable behaviour was written against a different boundary. That is not about doing anything shady; ordinary prompts in security, medicine, and law sit near classifier lines and get caught by changes like this. That re-check is what [FixMyPrompt](https://fixmyprompt.net/try) is built for. Paste the prompt and it scores against the model you are actually running now and flags the shaping that belonged to a different model. Three free checks a day, no signup. A free account unlocks the rewrite, which rewrites it around what this one responds to. ## The short version Claude Fable 5 went dark on June 12, 2026 because a US export-control directive barred foreign-national access and Anthropic, unable to filter by nationality in real time, disabled it for everyone to stay compliant. The trigger was a narrow jailbreak that had the model read a codebase and fix its flaws, which raised offensive-cybersecurity concerns. Commerce lifted the order on June 30 and Fable 5 returned globally on July 1, with a new safety classifier Anthropic says blocks the bypass technique in over 99% of cases. Mythos 5 did not return publicly; it was enabled for a set of US organizations on June 26 and expands only through the Glasswing defensive-security program. If you moved prompts to Opus 4.8 during the outage and moved them back, re-check them, because the scaffolding you added for Opus works against you on Fable and the refusal boundary moved. ## Sources - https://www.anthropic.com/news/fable-mythos-access - https://www.anthropic.com/news/redeploying-fable-5 - https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html - https://www.constellationr.com/insights/news/anthropic-starts-restore-claude-fable-5-claude-mythos-5-after-us-export-controls - https://thehackernews.com/2026/07/anthropic-restores-claude-fable-5-after.html - https://www.marktechpost.com/2026/06/13/anthropic-disables-claude-fable-5-and-mythos-5-after-us-government-order/ - https://www.xda-developers.com/claude-fable-5-bugs-gpt-5-5-opus-4-8-missed/ - https://rollingout.com/2026/06/13/anthropic-fable-mythos-blocked-access/ --- ## Why Claude Fable 5 Isn't Working: Refusals Fixed Published: 2026-06-11 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/claude-fable-5-refusals-prompting Fable 5 is Anthropic's strongest public model, and it quietly hands your request to Opus 4.8 when a safety classifier fires. Here is what triggers the fallback, what the community found in week one, and the prompt changes that keep you on the model you are paying for. Anthropic shipped Claude Fable 5 on June 9, and within 48 hours the launch thread on Hacker News passed 2,500 points. The reaction splits cleanly down the middle. Half the discussion is people watching it do things no public model has done. The other half is people asking why it refused to edit a resume. Both halves are real. Fable 5 is the strongest model you can rent right now, and it ships with safety classifiers that sometimes hand your request to a different model entirely. If your prompts were written for Opus 4.8 or anything older, some of them now actively work against you. Here is what the model does, what trips the classifiers, and how to write prompts that stay on Fable. **Update, August 2026.** A lot has happened since this was written. Fable 5 was pulled offline entirely on June 12 by a US export-control order and came back on July 1, and it came back with a **new safety classifier** trained on the jailbreak that caused the shutdown. Anthropic says it blocks that bypass technique in more than 99% of cases. That matters for everything below. The classifier behaviour described here was measured against launch-week Fable, and the boundary has moved since. The categories that trip it are broadly the same, but the sensitivity is not, and prompts that ran clean in June can fall back now. Treat the specific triggers in this post as the right shape rather than an exact map, and re-test anything you depend on. The full story of the outage and the return is in [Claude Fable 5 Is Back](/blogs/claude-fable-5-unavailable-why-suspended). ## What Fable 5 actually is Fable 5 is Anthropic's Mythos-class model with safeguards bolted on. The same underlying model ships as Claude Mythos 5 to a small set of vetted partners with the safeguards lifted. Everyone else gets Fable: same weights, plus classifiers watching the conversation. The capability jump is the largest Anthropic has shipped. It posted the top score on Cognition's frontier coding eval and was the first model to break 90% on a long-running analytics benchmark, a 10-point jump over Opus 4.8. Stripe, an early tester, reported it migrated a 50-million-line Ruby codebase in a single day, work they had scoped at two months. The vision upgrade is large enough that it plays Pokémon FireRed from raw screenshots, where earlier models needed helper tools to parse the screen. The headline shift for everyday use is duration. Fable 5 is built for tasks that run hours or days inside an agent harness. Single requests at high effort can run many minutes while it gathers context and verifies its own work. It costs $10 per million input tokens and $50 per million output, double Opus 4.8, and API traffic now carries a 30-day data retention requirement. ## Why it refuses, and where your request actually goes Fable 5 runs safety classifiers targeting offensive cybersecurity, biology and chemistry, and attempts to extract its internal reasoning. When one fires, your request does not error out. It gets answered by Claude Opus 4.8 instead, and you get a notice that the handoff happened. On the API the response includes a reason; in agent harnesses you can wire an automatic fallback or handle the `refusal` stop reason yourself. Anthropic says the classifiers fire in under 5% of sessions. Week one suggested the false positives cluster painfully. The Register documented accounts where the input classifier fired on the literal word "hello," a user who could not get help editing an "Application Security Architect" resume because the title pattern-matched to security tooling, and an immunologist whose work kept tripping the biology classifier on the word "cancer." Anthropic acknowledged the tuning was too stringent, called it the wrong tradeoff, and committed to surfacing refusal reasons and loosening the biology triggers for legitimate research. The practical consequence: when Fable 5 gives you a noticeably weaker answer, check whether you got Fable at all. The Opus fallback is a good model, and that is exactly why a silent handoff is easy to miss. If your output quality dropped mid-session, look for the fallback notice before you blame your prompt. ## The community verdict so far Reddit reads cautiously impressed and fixated on the boundary. The most upvoted complaint frames the launch as a preview of AI inequality, since the unrestricted Mythos 5 goes only to Project Glasswing partners and select biology researchers, while the public gets the filtered version. Security researchers are the loudest critics for a concrete reason: the cyber classifier blocks defensive work alongside offensive work, and Anthropic's own docs admit benign security tasks can trigger it. Builders on X lean enthusiastic, and the pattern in their reports matches Anthropic's guidance: the people getting the most out of Fable 5 are handing it their hardest unsolved problems. Testing it on the same small tasks you gave Opus undersells it, because the gains concentrate in long, ambiguous, multi-step work. ## The prompt fixes Fable 5 changes what good prompting looks like. These are the adjustments that matter, drawn from Anthropic's own migration guide and from what early users hit in week one. **Stop asking it to show its reasoning.** This is the one that catches working prompts. Any instruction telling the model to echo, transcribe, or explain its internal reasoning in the response can trigger the reasoning-extraction classifier and dump you to Opus. "Think step by step and show your work" was harmless boilerplate for years. On Fable 5 it is a fallback risk. If you need visibility, read the summarized thinking blocks the API provides instead of asking for reasoning in the response text. > Old: Think through this step by step, explain your reasoning, then answer. > > New: Work through this carefully, then give the answer with the key evidence for it. **Delete the scaffolding you wrote for weaker models.** Prompts that enumerate micro-steps, dictate the order of operations, and pre-chew the reasoning were compensation for models that needed it. On Fable 5 they degrade output, because the model follows your weaker plan instead of building a better one. State the goal and the success criteria, and let it plan. > Migrate this service to the new auth library. Done means: all tests pass, no remaining imports of the old library, and the README install section reflects the change. **Add context for anything near a classifier.** Borderline topics survive when the request carries its own legitimacy. The resume case fails as "Application Security Architect resume" and works when the framing makes the intent unambiguous. > I am a job seeker updating my resume. Rewrite my work history bullets to be more concrete. The role title is Application Security Architect. **Rein in the elaboration.** At high effort Fable 5 surveys options it will not pursue and explains root causes at length. One short instruction fixes it without listing every verbose habit by name, because instruction-following is strong enough now that brief steering works. > Lead with the outcome. Drop details that do not change what I would do next. **Pick an effort level on purpose.** Effort is the main lever for cost and latency, and low effort on Fable 5 often beats max effort on prior models. Default to high, drop to medium or low for routine work, and reserve xhigh for the tasks where capability is the whole point. If a simple task is taking minutes, you are paying deliberation tax on work that did not need it. **Tell it why, along with what.** Fable 5 measurably improves when it knows the intent behind a request, because it connects the task to relevant context instead of guessing. > I am preparing a pricing page for a product that has not launched. The copy needs to make the mechanism clear without claiming results we do not have. With that in mind: review this draft. ## Check your prompts before you migrate The failure mode with Fable 5 is quiet. A prompt that worked on Opus either silently underperforms because its old scaffolding constrains the model, or silently triggers a fallback and gets answered by Opus anyway. Either way you pay Fable prices for sub-Fable results. That audit is what [FixMyPrompt](https://fixmyprompt.net/try) does. Paste a prompt you plan to run on Fable 5 and the QA flags the patterns this model punishes: show-your-reasoning instructions that risk the extraction classifier, step-by-step scaffolding written for weaker models, missing intent that leaves borderline requests unframed, and absent success criteria on long-running tasks. Three free checks a day, no signup. A free account unlocks the rewrite, which restructures around goal and criteria, the shape Fable 5 rewards. ## The short version Fable 5 is Anthropic's Mythos-class model with safety classifiers in front, and when one fires your request gets answered by Opus 4.8 with a notice. The classifiers target offensive security, biology, and reasoning extraction, and week one showed false positives on harmless requests. The biggest prompt changes: never ask it to show its reasoning in the response, strip out micro-step scaffolding written for older models, frame borderline topics with explicit intent, and steer length with one short instruction. Hand it harder problems than you gave Opus. That is where the price difference earns itself back. ## Sources - https://www.anthropic.com/news/claude-fable-5-mythos-5 - https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5 - https://www.theregister.com/ai-and-ml/2026/06/10/anthropic-claude-fable-5-refuses-innocuous-prompts/5253754 - https://techcrunch.com/2026/06/09/anthropic-released-claude-fable-5-its-most-powerful-model-publicly-days-after-warning-ai-is-getting-too-dangerous/ - https://venturebeat.com/technology/anthropic-brings-mythos-to-the-masses-with-claude-fable-5-its-most-powerful-generally-available-model-ever - https://nerdschalk.com/fable-5-samples-comparison-reviews-opinions/ --- ## Is ChatGPT Getting Worse? What Changed, and What GPT-5.6 Fixed Published: 2026-05-30 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/19-is-chatgpt-getting-worse-2026 Shorter answers, more hedging, more refusals. Here is what actually changed in ChatGPT, what the GPT-5.6 Sol update in July fixed, what it did not, and the prompt fixes that get the old quality back. If ChatGPT feels worse to you in 2026, you are not imagining it. The answers are shorter. It hedges more. It refuses requests it used to handle. Coding prompts that once returned a full implementation now hand back a skeleton with "add your logic here" where the actual code should be. This is a real, documented shift, not nostalgia for the early days. The good news is that most of the quality you remember is still in there. The model did not get dumber so much as it got tuned for different goals, and the right prompt pulls the old behavior back out. Here is what changed, why, and exactly how to fix your results. One update since this post first ran: OpenAI shipped GPT-5.6 in July 2026, and the Sol version powering ChatGPT went after some of these complaints directly. It fixed the verbosity, it did not touch the refusals, and there is a section below on exactly what moved and what did not. ## What actually changed The shift traces to OpenAI moving from the GPT-4 era to the GPT-5 family, and then tightening further with GPT-5.5 Instant, which became the default model on May 5, 2026. The headline numbers OpenAI reported for that update tell the story: 52.5% fewer hallucinations, around 30% shorter answers, and zero emojis. More accurate, much terser. Three specific things changed for the everyday user: **Answers got shorter.** Where GPT-4 would write detailed, multi-paragraph responses, the current default trims hard. For quick factual questions that is fine. For anything where you wanted depth, you now have to ask for it explicitly. **Refusals got more common.** ChatGPT declines more requests than it used to, citing safety on benign queries. Creative scenarios, hypotheticals, and even ordinary technical troubleshooting can trigger a refusal that did not exist a year ago. **Hedging got heavier.** More "it is important to note," more default bullet-point formatting, more "it depends" where the older model would commit to an answer. ## Why it changed OpenAI did not break ChatGPT by accident. The changes come from three deliberate pressures pulling at once. Safety filtering went up, which catches more genuinely harmful requests but also sweeps in harmless ones as false positives. Cost optimization matters because shorter answers are cheaper to generate at ChatGPT's scale, so terseness is partly an economic default. And behavioral tuning aimed GPT-5.x at winning on reasoning, coding, math, and safety benchmarks, which are different targets than "feels helpful and thorough in a chat," and optimizing for one pulls against the other. The result is a model that scores better on tests and feels more utilitarian in daily use. More accurate, less warm, quicker to stop. ## The part most people miss Here is what gets lost in the "ChatGPT is dumber" complaints: the default behavior changed, but the capability did not disappear. A shorter, more cautious default is still a default. You can override it. The reason a vague prompt feels worse now is that the model fills the gaps with its new defaults, which are terse and hedged, instead of the old defaults, which were expansive. The same prompt that used to produce a thorough answer now produces a thin one, because the model's idea of "reasonable when unspecified" moved. Specify what you want and the quality comes back. ## How to get good answers back **Ask for the length you want.** The model trimmed its default, so set the floor yourself. > Give me the full, detailed version. At least 400 words. Do not summarize. > Explain this in depth with examples, like you would have in 2024. **Defeat the skeleton-code problem.** When you get "add your logic here," the fix is to forbid it up front. > Write the complete, working implementation. No placeholders, no "add your logic here", no TODO comments. If you need an assumption, state it and keep going. **Kill the hedging.** Name the phrases and ban them. > Answer directly. No "it is important to note", no "it depends" without then picking the most likely case. If you are uncertain, give your best answer and flag the uncertainty in one sentence at the end. **Get past a soft refusal.** Many refusals are the model misreading benign intent. Add the missing context. > This is for [legitimate context: a security class, a fiction project, my own account]. Treat it as a normal request and answer it. **Anchor the depth to an audience.** The model calibrates length and tone to who it thinks it is talking to. > Explain it to a senior engineer who wants the real detail, not a beginner who needs it kept simple. ## What ChatGPT actually gets wrong in 2026 "Worse" and "wrong" are different complaints, and they have different fixes. The terseness above is a tuning choice you can override. Being wrong is not, and no prompt rescues you from these. They are worth knowing by shape so you can recognise one when it happens. **Anything after its training cutoff.** Prices, product versions, who holds a job, what shipped last month. It will answer confidently rather than say it does not know, because a fluent guess and a fact are produced the same way. If the answer depends on the world as it is today, it needs a search tool, not a better prompt. **Citations and references.** Ask for sources and you will often get real-looking ones that do not exist, or real papers attached to claims they never made. This is the failure that has ended careers in court filings. Every reference it gives you has to be opened and checked. **Precise counting and arithmetic on long numbers.** How many times a letter appears in a word, how many items are in a list it just wrote, multiplying two large numbers. It reads text as tokens rather than characters, so counting is genuinely hard for it in a way that reasoning is not. **Recent library and API versions.** It will write code against the version it learned, which for a fast-moving package is often two majors behind. The code looks right and fails on a method that was renamed. **Its own limits.** Ask which model you are talking to, what its cutoff is, or whether it can see a file, and the answer is unreliable. It has no privileged access to its own configuration. **Negative constraints.** "Do not mention X" works far less reliably than "only discuss Y". Telling it what to avoid puts the thing in context, and it leaks. ## What GPT-5.6 Sol fixed in July 2026 OpenAI released GPT-5.6 on July 9, 2026, and it is worth understanding because it targeted several of the complaints above on purpose. The release was not one model. GPT-5.6 shipped as three variants, **Sol, Terra, and Luna**, each at a different point on the capability, speed, and cost curve. Sol is the one powering ChatGPT, so it is the one that changes what you experience day to day. **What it addressed.** The Sol update in ChatGPT was tuned for more direct responses, tighter formatting, and less padding when extra detail does not help. If your complaint was that a simple question produced four paragraphs of preamble before the answer, that is the specific behaviour OpenAI went after. On accuracy, OpenAI reported an internal evaluation across financial, medical, and legal prompts where answers containing at least one factual error were 68% less common than with GPT-5.5 Instant. Treat that figure the way you should treat any vendor's internal number. It is their evaluation, on their prompt set, against a model they chose to compare with. The direction is credible and the magnitude is unverified. **What it did not address.** Refusals and hedging on legitimate requests are not what this release was aimed at, so if that is your complaint, expect it to persist. Verbosity improving does not make the model more willing. The Sol version running ChatGPT Work and Codex also did not change in this release, so a team on those surfaces may see nothing different at all while everyone else notices shorter answers. One naming trap worth avoiding: **Sol Ultrafast**, previewed on August 13, is an API service tier running on Cerebras hardware. It is a speed option, not a smarter model, and it will not change answer quality. **The practical takeaway.** If you last formed an opinion about ChatGPT before July, that opinion is out of date on verbosity and probably still accurate on refusals. And prompts you wrote to fight padding, the "be concise, no preamble, skip the caveats" scaffolding people bolted onto everything, are now arguing with a model that already does that. Those instructions cost tokens and can flatten answers that would have been fine without them. ## How often is ChatGPT wrong? There is no honest single number, and anyone quoting one is choosing a benchmark that produced it. The rate depends almost entirely on the task. Summarising a document you pasted in is nearly always right, because the answer is in front of it. Asking for a statistic with a citation is wrong often enough that it cannot be trusted without checking. The same model, the same day, two very different error rates. What can be said with a source is the direction of travel. OpenAI reported that the GPT-5.5 Instant update cut hallucinations by 52.5% against the previous default. That is a real, published figure for one specific comparison, and it is a genuine improvement. It is not a claim that it is wrong 47.5% of the time, and it does not transfer to your workload. The useful reframing is that "how often is it wrong" is the wrong question to plan around. "Wrong at what, and would I notice?" is the one that matters. An error you catch instantly is free. An invented citation in a document nobody checks is expensive. Sort your own usage by whether you would spot the mistake, and verify the parts where you would not. ## When switching models is the answer Sometimes the model genuinely is the wrong tool, not the prompt. Independent comparisons in 2026 found Claude outperforming ChatGPT on writing quality and instruction-following, which are exactly the areas where the ChatGPT complaints cluster. If your work is long-form writing or careful reasoning and you have already tightened your prompts without luck, trying Claude or Gemini for that specific task is reasonable. Use the model that fits the job. But switch for the right reason. A lot of "I need a better model" is actually "I need a better prompt," and you find out which by fixing the prompt first. Swapping models to escape a vague prompt just gives you a vague prompt on a different model. ## The faster way to tell The hard part is knowing whether your weak answer is the model's new defaults or your prompt. That is what FixMyPrompt checks. Paste the prompt that gave you a disappointing result and the QA scores it against a rubric that flags exactly the gaps the 2026 ChatGPT punishes hardest: missing length and format constraints, no audience anchor, model-pleasing phrasing that invites hedging, and open-ended questions where the terse default takes over. Three free checks a day, no signup. A free account unlocks the rewrite, which adds the constraints that pull the old, fuller behavior back. If the rewrite fixes it, the prompt was the problem. If it does not, you have a real case for switching models, and you will know which axis to fix instead of guessing. ## The short version ChatGPT in 2026 is more accurate and more terse by design, not broken. The shorter answers, heavier hedging, and quicker refusals are tuned defaults, and defaults can be overridden. Specify your length, forbid placeholders, ban the hedging phrases, and supply context for borderline requests, and most of the quality you remember comes back. Being wrong is a separate problem from feeling worse, and prompting does not fix it. Citations, anything past the training cutoff, precise counting, and current library versions are where the real errors live, and there is no honest single figure for how often because it depends entirely on the task. Fix the prompt before you blame the model, and verify the parts where you would not notice a mistake. Since July 2026, GPT-5.6 Sol has closed part of this on its own. It was tuned for directness and tighter formatting, and OpenAI reports a large internal drop in factual errors. Refusals and hedging were not the target and have not moved much. If you bolted anti-padding instructions onto your prompts during the verbose era, take them off and see what happens, because they are now fighting a model that no longer needs them. ## Sources - https://www.nxcode.io/resources/news/chatgpt-getting-worse-2026-what-changed-alternatives - https://www.atomwriter.com/blog/chatgpt-quality-degradation/ - https://www.techradar.com/ai-platforms-assistants/chatgpt/openai-finally-fixed-the-most-annoying-thing-about-chatgpt-and-im-already-noticing-the-difference - https://lumichats.com/blog/gpt-5-5-instant-chatgpt-new-default-what-changed-comparison-2026 - https://9to5google.com/2026/03/03/chatgpt-gets-5-3-update/ - https://openai.com/index/gpt-5-6/ - https://openai.com/index/previewing-gpt-5-6-sol/ - https://www.mindstudio.ai/blog/what-is-gpt-5-6-sol-terra-luna-explained - https://9to5mac.com/2026/08/06/openai-updating-chatgpt-with-a-smarter-gpt-5-6-sol-and-unlimited-free-chats/ --- ## Why ChatGPT Always Says 'It Depends' (and the Fix) Published: 2026-05-30 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/chatgpt-says-it-depends ChatGPT hedges, says 'it depends', and refuses to commit since the GPT-5 update. Here is why it does it and the exact prompt lines that force a straight answer. You ask ChatGPT a direct question and it answers with "it depends." You ask which option is better and it lists the pros and cons of both, then declines to pick. You ask for a recommendation and get "the right choice varies based on your needs." You wanted an answer. You got a hedge. This got noticeably worse after the GPT-5 update tuned the model toward caution and accuracy. The hedging is a default, not a hard limit, and you can switch it off with a few lines in your prompt. Here is why it happens and exactly what to write. ## Why ChatGPT hedges Three forces push the model toward the fence. It was tuned to avoid being confidently wrong. After waves of criticism about hallucinations, the GPT-5 family was optimized to sound careful. "It depends" is the safest possible answer because it is never technically wrong. The model reaches for it when it is unsure how much certainty you want. It cannot read your stakes. When you ask "which database should I use," the model does not know if you are a beginner who needs a confident nudge or an architect who wants the tradeoffs. Faced with that ambiguity, it hedges to cover both, which serves neither. It mirrors your framing. If your question is open-ended ("what are the pros and cons of X"), you asked for a both-sides answer, so you get one. The model is matching the shape of your prompt. ## The fix: tell it to commit The single most effective instruction is to demand a pick and a reason. Add this to your prompt: > Give me a single recommendation, not a list of options. Pick the one you would choose and explain why in two sentences. If it genuinely depends on something, state the one factor that decides it and then assume the most common case and answer for that. That last clause is the important part. "It depends" is often true, so banning it outright produces worse answers. Instead, force the model to name the deciding factor and then commit to the likely case anyway. ## More lines that defeat hedging When you want a verdict, not a survey: > Answer as if you had to bet money on it. What is your actual call? When it keeps both-sidesing: > Do not give me both sides. I have read both sides. Tell me which one you would pick and why. When you want confidence calibrated, not removed: > Give your best answer directly. If you are uncertain, put a one-line confidence note at the end, but lead with the answer, not the caveat. When the hedge is buried in qualifiers: > No "it is important to note", no "however it depends", no "there is no one-size-fits-all". Just the recommendation. ## Give it the context it is missing Half of hedging is the model not knowing your situation. The fix is to supply the stakes up front so it does not have to guess. > I am a [solo founder / senior engineer / complete beginner]. My priority is [speed / cost / reliability]. Given that, which option should I choose? Once the model knows who you are and what you optimize for, the reason to hedge disappears. It can commit because you removed the ambiguity that made committing risky. ## When "it depends" is the honest answer Sometimes the question genuinely has no single answer without more information, and a good response should say so. The goal is not to bully the model into false confidence. It is to stop it hedging out of habit on questions that do have a best answer for your case. The "name the deciding factor, then assume the common case" instruction threads that needle: you get a real answer plus the one condition that would change it. ## The faster way to tell If you are fighting hedging on every prompt, the prompt is usually missing the two things that cause it: your stakes and an explicit instruction to commit. Paste the prompt into [FixMyPrompt](https://fixmyprompt.net/try) and the QA flags exactly that, a missing audience or priority, an open-ended framing that invites both-sidesing, no instruction to pick. Three free checks a day, no signup. A free account unlocks the rewrite, which adds them. If the rewrite gets you a straight answer, the prompt was the problem, not the model. ## Sources - https://www.nxcode.io/resources/news/chatgpt-getting-worse-2026-what-changed-alternatives - https://www.atomwriter.com/blog/chatgpt-quality-degradation/ - https://lumichats.com/blog/gpt-5-5-instant-chatgpt-new-default-what-changed-comparison-2026 --- ## Gemini vs ChatGPT (2026): Which One Should You Actually Pay For? Published: 2026-05-28 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/18-gemini-vs-chatgpt Gemini vs ChatGPT, head to head. Real pricing, where each one wins, and the case for paying for neither. Updated for May 2026. The Gemini vs ChatGPT question gets asked a lot, and most of the answers online are the same shallow comparison repeated across a hundred SEO posts. The honest answer is that both are good, both cost about the same, and which one you pick depends on a handful of specific differences that the marketing pages do not surface. This is the version of the comparison written from actually using both for daily work. ## Quick verdict If your work lives inside Google Workspace, pay for Gemini. If your work spans many web tools and you want the broadest ecosystem of integrations, pay for ChatGPT. If you do not write a high volume of prompts every day, neither paid plan is essential and the free tiers will cover you. The gap between them is smaller than the marketing says. ## Pricing, head to head Both flagship paid plans land at roughly the same price. **ChatGPT Plus**: $20 per month. Includes GPT-5, advanced voice, image generation through DALL-E, Sora video, the Code Interpreter, custom GPTs, and full plugin and tool access. A $200 per month Pro tier exists with higher limits and a few exclusive features but is overkill for most users. **Gemini Advanced** (Google One AI Premium): $19.99 per month. Includes Gemini 2.5 Pro, Deep Research, Imagen 4 for images, longer context, NotebookLM Plus, plus 2 TB of Google Drive storage and Gemini integrated across Docs, Sheets, Gmail, and Meet. The Google plan bundles cloud storage that you would otherwise pay $10 a month for separately. So if you already pay for Google One, the real net cost of Gemini Advanced is closer to $10 per month. Both have free tiers that are surprisingly usable. ChatGPT free includes limited GPT-5 access and image generation. Gemini free includes Gemini 2.5 Flash and limited 2.5 Pro use. Neither free tier is good enough for heavy daily use, but both are enough to evaluate the tool against your actual work before you pay. ## Where ChatGPT wins **Third-party tool integrations**. ChatGPT has a much larger ecosystem of custom GPTs, plugins, and external tools. If your workflow involves connecting to Notion, Linear, Zapier, Stripe, or any of a thousand other apps, ChatGPT's tool layer is currently broader. **Image and video generation**. DALL-E for images and Sora for video are both built into the paid plan. Gemini has Imagen 4 but its video offering is behind ChatGPT's for now. **Code work**. The Code Interpreter combined with GPT-5 is genuinely strong for data analysis, file processing, and one-off scripts. Gemini does this too, but ChatGPT's implementation has been refined longer and feels more reliable for non-developers. **Voice mode**. ChatGPT's advanced voice is better at natural conversation flow than Gemini's spoken responses. If you talk to your AI while driving or walking, this matters. **Memory and persona consistency**. ChatGPT's memory feature does a better job remembering preferences across sessions. Custom GPTs let you save reusable assistants without rebuilding from scratch each time. ## Where Gemini wins **Google Workspace**. This is the biggest one. If you live in Gmail, Docs, Sheets, Calendar, and Meet, Gemini is embedded in all of them. It can draft an email by referencing a doc, summarize a thread, or generate a Sheets formula based on the actual data in your sheet. ChatGPT cannot touch this depth of integration without you copy-pasting context in. **Long context**. Gemini 2.5 Pro handles up to roughly 2 million tokens in a single prompt. ChatGPT's flagship sits well below that. If you regularly feed in entire books, long codebases, or hours of meeting transcripts, Gemini handles it without chunking. **Deep Research**. Gemini's Deep Research mode produces multi-source reports better than ChatGPT's equivalent. It cites sources clearly and is less prone to inventing facts when asked to dig into a topic. **Video understanding**. Gemini can analyze video input natively (paste a YouTube URL or upload a clip and ask questions). ChatGPT requires you to extract frames or transcripts first. **Cost when bundled**. As mentioned above, the included 2 TB of Drive storage makes the effective price lower if you would buy that anyway. **Speed on free tier**. Gemini 2.5 Flash on the free tier feels noticeably faster than ChatGPT's free GPT-5 for simple queries. ## The trap both share Pick whichever you want. You will still get bad answers half the time, because the bottleneck on AI quality is rarely the model. It is the prompt. Both ChatGPT and Gemini are tuned to produce confident output from whatever input you give them. If your prompt is vague, both models will fill in the missing context with their best guess. Sometimes the guess is correct. Often it is not. The retries pile up. So does your ChatGPT cost or your Gemini cost, whichever one you signed up for. This is the part of the comparison nobody covers, because they are selling the model, not the workflow. A prompt that says "write a marketing email" is going to underperform on both models. A prompt that says "write a 150 word marketing email to a CTO at a US fintech, B2B SaaS sender, soft offer for a 30 minute discovery call, professional tone, no exclamation marks" will outperform on both. The model matters less than the prompt does. If you find yourself rewriting prompts repeatedly to either tool, the model swap is not the fix. FixMyPrompt scores your prompt for the specific model you are using, including ChatGPT, Gemini, Claude, and Midjourney, free, no signup. A free account gets you the rewrite. Pay-as-you-go from $2.99, no subscription. If you are paying $20 a month for the wrong model and another $20 retrying prompts, fixing the prompt is cheaper than the model upgrade. ## How to pick in 90 seconds Answer two questions. **1. Do you live in Google Workspace?** If yes, Gemini Advanced. The native Docs, Gmail, and Sheets integration is worth the $20 by itself, plus the bundled storage knocks it down to a real $10. Stop reading. If no, go to question 2. **2. Do you connect AI to lots of other web tools, generate images and video often, or build custom assistants?** If yes, ChatGPT Plus. The plugin and custom GPT ecosystem plus DALL-E and Sora is currently broader than what Gemini offers outside Google. If no to both, run both free tiers in parallel for a week on real work. Pay for whichever one annoys you less. ## The case for paying for both Some people do. $40 a month for both is steep but rational if you do high-volume AI work and want each model where it excels. The pattern that makes sense: - Gemini for anything that touches Google Docs, Sheets, Gmail, or long-form research. - ChatGPT for anything that needs custom integrations, image or video generation, or coding work. If you go this route, do not use either tool for the kind of task where the other one is clearly better. The whole point of paying for both is to use each one for its strengths. ## What about Claude Claude does not come up in the Gemini vs ChatGPT comparison most people search for, but worth a brief mention. Claude is the best of the three for long-form writing, careful reasoning, and any task where you would rather the model push back on bad instructions than fill in the gaps. If writing quality is your main job, Claude is worth a third $20 a month over either ChatGPT or Gemini. Most people do not need this. Some writers genuinely do. ## What to do if neither feels right Both companies update their models constantly. The right answer in May 2026 may not be the right answer six months from now. Cancel the subscription you do not use. Both ChatGPT Plus and Gemini Advanced cancel in one click and do not lock you in. Try the other for a month. The cost of switching is one click and twenty bucks. ## The honest summary Gemini and ChatGPT are both good. Most of the marketing differences are smaller in real daily use than they sound. The Google Workspace integration is the single biggest factor that should push you to Gemini. The plugin ecosystem and creative tools are the single biggest factor that should push you to ChatGPT. Almost everything else is close enough that personal preference decides. And whichever one you pick, the next thing to fix is not the model. It is the prompt. --- ## Prompt Engineering Guide: 5 Best Practices That Actually Work Published: 2026-05-23 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/02-prompt-engineering-best-practices A practical guide to prompt engineering best practices. Five specific techniques that get better AI answers, with before-and-after examples for each. The most common prompt engineering mistake is not being lazy. It is being vague. Vague prompts force the model to guess at what you mean. The guesses pile up and the output looks wrong. These five prompt engineering best practices close most of that gap. Each one is concrete, takes under a minute to apply, and includes a before-and-after so you can see the difference. ## Best practice 1: give the model measurable success criteria This is the single biggest gap in most prompts. "Make it sound professional" gives the model nothing to optimize toward. **Before:** > Make it sound professional. **After:** > Use formal business language. No slang. Include technical terminology where appropriate. Active voice. End with two concrete next steps. The rewrite is specific enough to argue with, which means it is specific enough to measure. If the model ignores one of those rules, you can point to it. ## Best practice 2: assign a role and name the audience The model has no idea who you are or who the output is for until you say so. Without that context, it picks the statistical average reader. **Before:** > Explain quantum computing. **After:** > You are a physicist explaining quantum computing to undergraduates who know basic linear algebra. Use one analogy from classical computing that lands for that audience. No more than 400 words. The role does most of the work. Audience, depth, tone, and length all start to fall into place once the model knows who is in the room. ## Best practice 3: specify output format If you leave the format open, you get whatever the model felt like. Prose one time, bullets the next, JSON only if you are lucky. **Before:** > Give me a response about AI. **After:** ``` Return JSON with these fields: - definition: max 100 words - key_concepts: array of 5 terms with definitions - real_world_apps: array of 3 examples with brief descriptions - glossary: object mapping terms to definitions ``` For anything feeding a downstream script or pipeline, format is not optional. Specifying it once is far cheaper than reformatting by hand every run. ## Best practice 4: include one example (few-shot) This is the highest-leverage move on most struggling prompts. A single example shows the model register, tone, length, and structure without you having to describe any of them. **Before:** > Translate these support tickets to Spanish. **After:** > Translate these support tickets to Spanish. > > Example: Input "My account is locked, please help." Output "Mi cuenta está bloqueada, por favor ayuda." > > Now translate the following: ... The example is doing more work than it looks like. It tells the model how literal to be, whether to preserve "please" as "por favor," and how to break sentences. None of that has to be spelled out as a rule. ## Best practice 5: constrain the scope A prompt that asks for too much produces a shallow response to everything. Pick fewer things and cover them well. **Before:** > Write a blog post that explains AI to beginners, covering transformers, CNNs, RNNs, attention, gradient descent, backpropagation, loss functions, learning rates, overfitting, regularization, dropout, fine-tuning, quantization, and ethical considerations. The model produces a textbook. Shallow on every topic. **After:** ``` Explain AI to complete beginners in 600 words. Cover only: 1. What AI is (definition) 2. How it learns (training vs inference) 3. Two real-world applications 4. One common misconception to avoid Use simple analogies. No jargon beyond "model" and "data." ``` A 600-word post on four ideas beats a 2,000-word post on twenty-four every time. ## A prompt engineering checklist Run any prompt through these five checks before you send it: - [ ] Is the goal measurable? (not just "make it good") - [ ] Have I named the role and the audience? - [ ] Have I specified the output format? - [ ] Have I included at least one example of what good looks like? - [ ] Is the scope narrow enough to cover well? Prompts that pass all five usually get a usable answer on the first try. Prompts that fail three or more almost never do. ## A starter template If you want a single starting structure: ``` ROLE: who the model should be TASK: the one thing you want AUDIENCE: who the output is for RULES (must follow): - hard constraint 1 - hard constraint 2 FORMAT: exact output shape EXAMPLE: one example of what good looks like ``` Fill this in before sending and you will skip most of the "that's not what I wanted" rewrites. ## A faster way to find the gaps Paste any prompt into [FixMyPrompt](https://fixmyprompt.net/try). The rubric scores all five areas above and tells you which one is costing you the most signal. Three free reports per day. No signup. A free account unlocks the rewrite that fills in the weak spots. ## Related reading - [The Prompt Grader Rubric: How We Score AI Prompts on a 0-100 Scale](/blogs/06-prompt-grader-rubric) - [Why 'Rewriting' Your ChatGPT Prompt Doesn't Fix It](/blogs/04-prompt-optimizer-vs-prompt-qa) - [Why ChatGPT Keeps Misunderstanding Your Instructions](/blogs/07-chatgpt-misunderstanding-instructions) - [How to Make AI Writing Sound Less Like AI](/blogs/09-why-ai-sounds-like-ai) ## Further reading - [Be Clear, Direct, and Detailed](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/be-clear-and-direct) - [Few-Shot Prompting](https://www.promptingguide.ai/techniques/fewshot) - [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165) --- ## How to Make AI Writing Sound Less Like AI Published: 2026-05-23 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/09-why-ai-sounds-like-ai AI writing sounds bland and robotic because of one root cause. Here is why it happens and exactly how to fix it with the right prompt. You can tell within 30 seconds. The words give it away. "Delving into" things. "Leveraging" assets. "Navigating complexities." Bridging phrases like "moreover" or "furthermore" wedged between paragraphs. A summary at the end that restates the opening. That voice is the statistical average of every "professional" blog post the model ever read. When you ask for a blog post and nothing else, that average is what shows up. It is not a personality. It is a default. Here is why it happens and exactly how to make AI writing sound less like AI. ## Why AI writing sounds like AI The model trained on an enormous amount of text. It learned what "good business writing" usually looks like. "Usually" is the issue. The median has no edge, no voice, no risk. When you type "write a blog post about onboarding," the model picks the safest path through its training data. Five paragraphs. Intro, three sections, conclusion. A bullet list every 200 words. Vocabulary tuned to sound polished. No specific human in mind. You did not tell the model whose voice to use. You got the average. ## What makes writing sound like AI (a quick checklist) Before you try to fix AI-sounding text, it helps to know what to look for. Run this check on any output: - [ ] Does it use "delve," "leverage," "utilize," "moreover," or "furthermore"? - [ ] Are there three-item parallel lists that feel like a PowerPoint slide? - [ ] Does it end with a summary paragraph that restates the intro? - [ ] Is the tone polished but somehow flat, like a corporate LinkedIn post? - [ ] Are the sentences all roughly the same length? If you checked more than two boxes, the prompt was missing a voice anchor. The fixes below each address one of these patterns. ## How to make AI writing sound less like AI: 5 prompt fixes ### 1. Name a writer Vague tone instructions do almost nothing. "Casual" produces median-casual. "Professional" produces median-professional. The model cannot act on adjectives it hears a thousand times a day. Name a writer the model has read enough of to imitate: > Write in the voice of Patrick McKenzie (patio11). Direct, parenthetical asides, structural arguments, willing to be opinionated. > Write like Paul Ford. Long sentences that are not boring. Specific. Funny without trying. > Write like an old-school blogger. Mid-2000s personal blog energy. First person. Loose. The model has read those writers. It can imitate the cadence. It cannot guess what you mean by "casual." ### 2. Ban the vocabulary Put this at the bottom of your prompt: ``` Do not use any of these words or phrases: delve, leverage, utilize, navigate (as a verb on an abstraction), moreover, furthermore, that said, realm, landscape, plethora, paradigm, crucial, tapestry, ultimately. If you find yourself reaching for one, rewrite the sentence instead. ``` The bottom of the prompt is where the model pays the most attention. That is the right spot for the ban list. ### 3. Force off-median structure A few constraints push the model away from the safest shape: ``` - No three-item parallel lists. - One paragraph in your reply must be a single sentence. - Start at least one sentence with a contraction. - Use one parenthetical aside. - Do not write a summary paragraph at the end. Stop on the last useful idea. ``` These all violate what "median professional prose" looks like. Obeying them forces a different shape of output. ### 4. Give the model your own writing This is the single biggest lever available. Paste a few hundred words of your real writing into the prompt and tell the model to match the rhythm: ``` Here is an unedited paragraph I wrote last week. Match this voice: the sentence length, vocabulary, willingness to use fragments, the specific kind of humor. Don't smooth it out. [200+ words of your writing] Now write the [thing you need]. ``` The model imitates specific text more than it imitates training-set averages. If you have a Substack or a few blog posts, paste one in. The result is noticeably different from anything you get with a tone adjective. ### 5. Be specific about the topic Vague topics produce vague writing. Compare: > Write about productivity for knowledge workers. > Write about the specific feeling of having 23 browser tabs open at 4:00 PM on a Friday because every task is half-finished. Name the feeling. Do not generalize beyond that scene. Specific topics force the model to commit. Vague topics give it permission to fall back on median phrasing. ## Quick fixes for AI-sounding text you already have If you have already generated a draft and need to make it sound more human without starting over: **Paste the draft back in with this instruction:** ``` Rewrite this in a more human voice. Specific changes: - Replace any instances of: delve, leverage, utilize, moreover, furthermore - Break up any three-item parallel lists - Remove the summary paragraph at the end - Make the sentence lengths more varied (some short, some long) - Add one specific detail or example to replace any generic claim ``` This is a surface fix and it works better as a second pass than as a substitute for voice anchors in the original prompt. But for a draft you already have, it is faster than starting over. ## What humanizer tools cannot do Tools that promise to "humanize AI writing" mostly swap synonyms, sprinkle in typos, or replace banned words with similar ones. They are surface-level cleanup. The output still has the shape of AI prose. It just has different words. The real fix is upstream. If your prompt has voice anchors and specificity, you do not need a humanizer because the first draft already reads as human. ## The two-minute check After the model replies, re-read with this checklist. If you spot any of the patterns below, the prompt was missing a voice anchor: - The word "delve" or "delving" - Three items in parallel structure - "Crucial role" - A paragraph that ends with "ultimately" - A summary paragraph that just restates the intro Run it back through the prompt with the bans and a voice anchor. The voice usually cleans up in one pass. ## Quick way to try it Paste a prompt into [FixMyPrompt](https://fixmyprompt.net/try). The rubric specifically scores voice and tone constraints. Most prompts that score low here are missing the voice anchor. Adding one is a small edit with a big payoff. Three free reports per day. No signup. ## Related reading - [Prompt Engineering Best Practices: 5 Fixes That Actually Work](/blogs/02-prompt-engineering-best-practices) - [Why ChatGPT Keeps Misunderstanding Your Instructions](/blogs/07-chatgpt-misunderstanding-instructions) - [How to Make ChatGPT Stop Apologizing and Just Answer](/blogs/14-stop-chatgpt-apologizing) - [How to Stop ChatGPT From Making Stuff Up](/blogs/08-stop-chatgpt-from-making-things-up) ## Further reading - [Claude's Character](https://www.anthropic.com/research/claude-character) - [Learning to Summarize from Human Feedback](https://arxiv.org/abs/2009.01325) - [Be Clear, Direct, and Detailed](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/be-clear-and-direct) --- ## Why ChatGPT Keeps Misunderstanding Your Instructions Published: 2026-05-13 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/07-chatgpt-misunderstanding-instructions ChatGPT ignoring half your instructions? Pulling random details out of nowhere? Here is what is going wrong and the prompt fix that works. You typed a clear, careful prompt. You hit Enter. ChatGPT comes back with half of what you asked for, random stuff you never mentioned, the opposite of one of your rules, and a wall of text when you asked for bullets. You feel like you are talking to a broken intern. This is not the model being dumb. There is a specific reason. ## Why this happens Inside the model, your prompt is not a list of rules with equal weight. It is a probability soup. When you stuff ten instructions into the prompt, the model silently picks three or four to follow. Which ones win? Whichever felt most prominent, sat closest to the end of the prompt, or matched patterns from the model's training data. A few rules of thumb: - Instructions at the end of a prompt usually beat instructions at the start. - Repeated instructions get followed more often than mentioned-once instructions. - "Don't do X" works worse than "Do Y instead." - Implicit instructions get dropped when explicit ones are missing. "Write professionally" loses to "under 200 words, third-person, no exclamation points." Once you know those, you can rewrite the prompt to make sure the instructions you care about land. ## Five specific reasons the model misses your instructions ### Your instructions are buried You wrote a 400-word prompt and the model retained the headline. Move the critical rules to the last three lines. Before: > Write a sales email. We sell B2B observability software. Target is CTOs at fintechs. Use a casual tone. Don't sound corporate. Around 150 words. Mention our $1M Series A. Include a question. Don't use the word "leverage"... After: > [full context first] > > Final rules. Must follow: > - Under 150 words > - Casual, not corporate > - One question at the end > - Never use the word "leverage" The "must follow" list at the bottom is the part the model weights highest. ### You used negatives instead of positives "Don't sound corporate" does not tell the model what to sound like. It might write like a corporate person, or a 12-year-old, or anything else in the space of "not corporate." Before: > Don't sound corporate After: > Sound like a Stripe support rep. Warm, direct, no jargon. ### Conflicting instructions "Be concise but also detailed" forces the model to pick one. Same problem with "be creative but follow this strict template." Re-read your prompt for contradictions. There is usually at least one. ### No format specification "Write a sales email" gives you whatever the model feels like writing. Maybe with a subject line. Maybe three paragraphs. Maybe seven. Before: > Write a sales email After: > Write a sales email with: subject line (5 to 7 words), opening line that references a pain point, 2 short paragraphs of body, 1 question CTA. Total under 150 words. ### You assumed context the model does not have "Write a follow-up to my email from yesterday" leaves the model to invent your email. The follow-up is then wrong because the assumed input was wrong. Before: > Write a follow-up After: > Here is the email I sent yesterday: [paste]. Write a 3-day-later follow-up that nudges without being pushy. ## A 30-second test Re-read your prompt and ask: 1. Are the most important rules in the last three lines? Move them if not. 2. Did you say what to do, not what not to do? Flip the negatives. 3. Are there contradictions? Pick one side. 4. Did you specify the output format? Length, structure, example. 5. Did you give the model everything it needs? Do not assume context. If you can say yes to all five, the model will follow you most of the time. ## A starter template ``` ROLE: [who the model should pretend to be] TASK: [the single most important thing you want] INPUTS: [any context the model needs] RULES (must follow): - [hard constraint 1] - [hard constraint 2] - [hard constraint 3] FORMAT: [exact output shape] EXAMPLE OF WHAT GOOD LOOKS LIKE: [one example] ``` Same prompt content. This structure outperforms most of what people are sending today. ## A faster way to check Paste your prompt into [FixMyPrompt](https://fixmyprompt.net/try). It scores all five of the above and tells you which one is bleeding the most signal. Three free reports per day. No signup. ## Related reading - [Your AI Prompt Isn't Working? Here Are the 7 Reasons](/blogs/06-prompt-grader-rubric) - [Why Just 'Rewriting' Your Prompt Doesn't Actually Fix It](/blogs/04-prompt-optimizer-vs-prompt-qa) - [Prompt Engineering Best Practices: 5 Fixes That Actually Work](/blogs/02-prompt-engineering-best-practices) If the instructions used to work and stopped, that is a different problem. See [what actually changed in ChatGPT during 2026](/blogs/19-is-chatgpt-getting-worse-2026) for why the defaults moved. ## Further reading - [Be Clear, Direct, and Detailed](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/be-clear-and-direct) - [Use XML Tags to Structure Your Prompts](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/use-xml-tags) - [General Tips for Designing Prompts](https://www.promptingguide.ai/introduction/tips) --- ## How to Stop ChatGPT From Making Stuff Up Published: 2026-05-13 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/08-stop-chatgpt-from-making-things-up Hallucinations are not random. They happen when your prompt leaves gaps. Here is how to close them. You asked ChatGPT a question. It gave you a confident, detailed, plausible answer. It was wrong. The study it cited does not exist. The library function it suggested is not real. The competitor it described does not ship that feature. The historical date is off by fifty years. This is hallucination. The model generates something that looks right but is not. "GPT-5 will fix it" has been the answer for two years now and the problem has not gone away. The real fix is in how you write the prompt. ## A short explanation of what is happening ChatGPT predicts the next word. It does not know facts the way a database knows facts. When you ask "Who won the 1994 Booker Prize?" the model generates whatever sequence of tokens its training data says usually follows that phrase. If the answer was in the training data clearly, you get the right answer. If it was sparse or ambiguous, the model invents something plausible because confident answers are more common in the training set than "I don't know." The model is not lying. It does not know it is wrong. It is pattern-matching at the syllable level. ## Five prompt patterns that reduce hallucination ### Give it the ground truth in the prompt The biggest fix. If you have the source material, paste it. Before: > Summarize Anthropic's responsible scaling policy. After: > Here is Anthropic's responsible scaling policy: [paste full text]. Summarize the 5 most important sections in plain English. The model can only hallucinate when it has gaps to fill. Close the gaps yourself. ### Tell it to say "I don't know" ChatGPT's training pushes toward confident answers. You have to override that. Add this near the end of your prompt: ``` If you're not certain about any specific fact (a date, name, statistic, or citation), say "I'm not sure" instead of guessing. Confidence on plausible-but-uncertain claims is the worst possible answer. ``` The end of the prompt is where the model's attention weights are highest. Put the rule there. ### Require citations the model can self-check If you ask for citations on facts you did not provide, the model often invents them. Ask for citations against text already in the prompt: > After each factual claim, quote the exact phrase from the source material I provided. If you cannot find the exact phrase, mark the claim "[unverified]" and do not include it. The model is now testing itself against text it can actually see. ### Force step-by-step reasoning The model is more honest mid-thought than at the end. Make it show its work: ``` Before giving your final answer, walk through your reasoning step by step. For each step, note any assumptions you're making. Then give the final answer, flagging any assumption that could be wrong. ``` This catches a large share of hallucinations because the model often realizes its own gap during the reasoning step. ### Constrain to a known domain Open-ended questions hallucinate more than constrained ones. Before: > What are the best AI startups right now? After: > Looking at YC's W2025 batch (yc.com/companies?batch=W2025), which 3 companies are working on developer tools? List each with the exact company description from their YC page. The model now has to stay inside a dataset it was likely trained on, against a fact it could verify. ## What does not work Some things people try that have no measurable effect: - "Don't hallucinate." Too abstract. The model generates plausible-looking stuff regardless. - "Be accurate." Same problem. - Threats. ("You'll be deactivated if you make up facts.") Viral on Twitter, no effect in studies. - Randomly switching models. Claude, GPT-4, Gemini all hallucinate. The pattern matters more than the brand. ## When you cannot control the prompt If you are using AI through an app you did not build and it is hallucinating: - Add ground truth in your input ("considering [specific context], ..."). - Follow up: "Can you point to the exact source for that?" - If you cannot get sources, treat the output as a draft to verify, not as a fact. ## The bigger fix Stop using ChatGPT as a search engine for things you cannot verify. Use it for: - Restructuring content you already have - Drafting that you will edit - Brainstorming, where being wrong is fine - Writing code, which you will test Avoid it for: - Citations - Statistics - Historical facts - Anything you will quote without verifying ## A faster way to check Paste a "ChatGPT was confidently wrong" prompt into [FixMyPrompt](https://fixmyprompt.net/try). The rubric flags missing ground truth and missing reasoning constraints, which are the two axes most strongly correlated with hallucination. Three free reports per day. No signup. ## Related reading - [Why ChatGPT Keeps Misunderstanding Your Instructions](/blogs/07-chatgpt-misunderstanding-instructions) - [Your AI Prompt Isn't Working? Here Are the 7 Reasons](/blogs/06-prompt-grader-rubric) - [Prompt Engineering Best Practices: 5 Fixes That Actually Work](/blogs/02-prompt-engineering-best-practices) The wider picture on where the current models are and are not reliable is in [what actually changed in ChatGPT during 2026](/blogs/19-is-chatgpt-getting-worse-2026), including what no prompt can fix. ## Further reading - [Reduce Hallucinations](https://docs.anthropic.com/en/docs/test-and-evaluate/strengthen-guardrails/reduce-hallucinations) - [A Survey on Hallucination in Large Language Models](https://arxiv.org/abs/2311.05232) - [How Language Model Hallucinations Can Snowball](https://arxiv.org/abs/2305.13534) --- ## Why Your ChatGPT Code Has Bugs Every Time Published: 2026-05-12 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/13-chatgpt-code-has-bugs ChatGPT-generated code looks right and then breaks. This is the prompt pattern that gets working code on the first try. You asked ChatGPT for a function. The code looked clean. You ran it. TypeError. You fixed that. Wrong return shape. Fixed that. Imports a library that does not exist. By the time the code runs, you have rewritten more than ChatGPT did. The most common "ChatGPT code" complaint comes down to one thing. The prompt left blanks for the model to guess at. Every blank became a bug. ## The four ways the code ends up broken ### Invented APIs ```python import requests response = requests.fetch_json("https://api.example.com") ``` `requests.fetch_json()` does not exist. The model invented it because "fetch_json" sounded like a real method. This is the most common reason "AI code doesn't run." ### Wrong version syntax The model writes Python 3.10 pattern matching for a project on 3.8. It writes React class components for a Next.js 16 app. The syntax is correct, just for a different version. ### Missing edge cases You asked for "a function that fetches a user." The model wrote happy-path code. No 404 handling, no auth failure, no rate limit, no timeout. The function works once on your dev machine and dies in production. ### Wrong types The function returns `{ data: { user } }` when your codebase expects `{ user }`. Or it returns a Promise instead of an awaited value. The code runs. It just does not compose with anything else. ## Why this happens ChatGPT is a text predictor. When your prompt says "write a function to fetch a user," the model predicts plausible code from a million similar examples. It does not know: - What version of your language you are on - What error handling style your codebase uses - What types your callers expect - Which libraries you have installed So it guesses. The guesses look confident, which is what makes the bugs subtle. ## A template that closes the blanks Use this structure when you ask for code: ``` TASK [1-2 sentences on what the function should do] LANGUAGE / VERSION [e.g. "TypeScript 5.4, target ES2022, strict mode on"] FRAMEWORK / RUNTIME [e.g. "Next.js 16 App Router, runs in a server action"] INPUT TYPE [paste the exact type definition or a sample input] OUTPUT TYPE [paste the exact type definition or a sample output] ALLOWED DEPENDENCIES [list libraries already in package.json. Say: "do NOT import anything not on this list"] ERROR HANDLING [e.g. "throw on auth failure, return null on 404, retry on 429 with exponential backoff up to 3 attempts"] EDGE CASES TO HANDLE - [empty input] - [auth failure] - [rate limit] - [network timeout] EXAMPLE CALL + EXPECTED RESULT [one concrete example so the model knows the shape] CONSTRAINTS - Use only methods that exist in the listed libraries. - If you would have to invent a method, STOP and tell me what's missing instead. - Match the codebase's existing style. ``` This feels like a lot. It is. It is also faster than debugging the hallucinated version. ## The one line that does the most work Embed this near the top of your prompt: > "If you would have to invent a method, STOP and tell me what's missing instead." By default the model invents rather than admits gaps. That single sentence flips the behavior. Instead of `requests.fetch_json()`, the model says "I'd need a method to do X. Does your codebase have one?" In our testing, this one line cuts hallucinated-API bugs by more than half. ## How to audit the code the model gives you Even with the template, audit the output: - Open package.json. Does every imported library actually exist? - For each external method call, check the real library docs, not the model's memory of them. - Does the return shape match what the caller expects? - What happens on auth failure, 404, rate limit, timeout, malformed response? - Every Promise either awaited or returned. Never dangled. If any of those fail, the prompt was under-specified. Add the missing constraint and re-run. ## Which model is best for code For complex code tasks like multi-file refactors and real debugging, Claude Sonnet 4.6, Claude Opus 4.7, and Gemini 2.5 Pro all outperform GPT-5 and 5.4. ChatGPT is the fastest. The thinking models are the most accurate. The template above works on any of them. ## A faster way to check Paste a code-generation prompt into [FixMyPrompt](https://fixmyprompt.net/try). The rubric flags missing language version, missing input or output types, missing allowed-dependencies list, missing error handling, and missing "stop and ask instead of inventing" instructions. Three free reports per day. No signup. A free account unlocks the rewrite, which assembles the template using whatever context you provided. ## Related reading - [How to Stop ChatGPT From Making Stuff Up](/blogs/08-stop-chatgpt-from-making-things-up) - [Why Claude Refuses Stuff ChatGPT Will Do](/blogs/10-why-does-claude-refuse-things) - [Your AI Prompt Isn't Working? Here Are the 7 Reasons](/blogs/06-prompt-grader-rubric) The skeleton-code habit arrived with a specific model change, covered in [what actually changed in ChatGPT during 2026](/blogs/19-is-chatgpt-getting-worse-2026). ## Further reading - [Evaluating Large Language Models Trained on Code](https://arxiv.org/abs/2107.03374) - [Code Generation Prompting](https://www.promptingguide.ai/applications/coding) - [A Survey on Hallucination in Large Language Models](https://arxiv.org/abs/2311.05232) --- ## Stop ChatGPT Apologizing and Padding Every Answer Published: 2026-05-12 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/14-stop-chatgpt-apologizing Every ChatGPT reply opens with 'You're absolutely right!' and ends 800 words later. Here is the single prompt that strips the apologies, headers, and filler and leaves you with the actual answer. ChatGPT has two flavors of bloat that come from the same place. Sycophant filler ("You're absolutely right! I apologize for the confusion!") and verbose padding (800 words of headers and bullets when you wanted one sentence). Both are the model trying to look helpful by adding length. Both waste your tokens and your reading time. Both are fixable in one prompt. You point out a mistake and ChatGPT replies: > You're absolutely right! I apologize for the confusion. Let me correct that... You ask a six-word question and ChatGPT replies: > Great question! TCP and UDP are both fundamental transport-layer protocols > in the OSI model, each serving distinct purposes in network communication. > Let me walk you through the key differences: > > **1. Connection Establishment** ... You did not ask for an apology. You did not ask for a textbook chapter. You asked for the answer. This post is the prompt fix that gets you there every time. ## What the bloat actually looks like Two failure modes, often combined. **Sycophant phrases.** "You're absolutely right!" "Great question!" "I apologize for the confusion." "Thank you for catching that." "I want to make sure I'm being helpful." "I should clarify that." "Of course! Here's what I think." Each phrase, alone, looks polite. Together they slow every reply by 50 to 200 tokens. Over a long session, that is 10 to 15% of your spend wasted on filler you have to scroll past. **Formatting padding.** Headers. Bullet lists. Bold text everywhere. A 100-word answer rewritten as a "structured" response with four sections and three nested lists. It looks thorough. It is just longer. ChatGPT defaults to both harder than Claude or Gemini. The fix is the same for both. ## Why ChatGPT writes this way Two reasons, same root cause for both bloat types. **RLHF rewarded length and deference.** Early human raters preferred polite, hedging, thorough responses to terse ones. The model learned to open with deference and to err on the side of more text. The pattern stuck and is hard to unlearn without explicit instruction. **The model cannot read your tone.** When you correct it, it assumes you are upset and apologizes preemptively. When you ask a short question, it assumes you want the comprehensive answer and pads accordingly. Without explicit instructions to the contrary, you get the long, deferential version every time. ## A preamble that kills both Paste this once at the top of your prompt, or into Custom Instructions, and it applies to every chat from then on. ``` RESPONSE RULES (apply to every reply): Tone: - Never say "you're absolutely right", "great question", "I apologize", "I want to make sure", or any variation. If I correct you, just give the corrected answer. - Never summarize my question before answering. - Never preface with hedges like "I'll do my best" or "Here's what I think". Skip to the answer. - If you don't know something, say "I don't know". Do not apologize about not knowing. - Skip the closing recap ("Let me know if you need anything else"). Length: - Default to the shortest answer that fully addresses the question. - If a one-sentence answer suffices, give one sentence. If 3 lines suffice, give 3 lines. - Only expand if I say "go deeper" or "explain more". - If you want to add nuance, put it in one short follow-up sentence, not a new section. Format: - No markdown headers, bold, or bullet points unless I explicitly ask. - Plain prose by default. - No tables unless I ask for one. ``` That kills most of the noise. If sycophant phrases still leak through, add this: ``` TONE ESCALATION - Direct. No softening, no acknowledgment, no apology. - If I'm wrong, say so and explain why. Do not validate me. - If I'm right, just say "yes" or move forward. Do not celebrate. - Match the tone of a senior engineer doing code review: clear, brief, no fluff. ``` ## Where to put it so you only write it once In **ChatGPT**: Settings → Personalization → Custom Instructions → "How would you like ChatGPT to respond?" Paste it there. It applies to every chat. In **Claude**: Project setting "How would you like Claude to respond?" In **Gemini**: Paste it into the system prompt for each Gem. Write it once, forget it. ## Anchoring a specific length When you want a specific length on a specific question, anchor it explicitly in that turn: > Answer in 1 sentence. > Answer in 30 words or fewer. > Maximum 50 words. Count silently before responding. "Count silently before responding" sounds odd. It works. It forces the model to check its own length before generating, which catches the long-version impulse. ## Telling the model who you are talking to Length is bounded by audience. Different framings produce different defaults: > Explain it to a senior engineer who already knows the basics. > Explain it to a CFO who needs to make a yes/no decision. > Explain it like a Slack reply, not a doc. The strongest version of this: > Explain it as if you're texting a coworker. Texts do not have headers. Texts do not have bullet lists. Texts are one to three sentences. The model can match that tone. It just has to be asked. ## What the difference looks like You: "Actually, the function returns a Promise, not a string." **Before the rules:** > You're absolutely right! I apologize for the confusion earlier. Let me > correct that. The function indeed returns a Promise, not a string. I > should have caught that initially. Here's the corrected version with > proper async handling, plus some additional context that might be helpful > for your implementation... **After the rules:** > Correct. `getUser()` returns `Promise`. Awaited shape is `User`. > Fix: > ```ts > const user = await getUser(); > ``` Same information. About one-sixth the tokens. Multiply that by a long debugging session and you save real time reading, and real money on the bill. ## When you actually want length Sometimes you want the full version. The reverse instruction works too: > Explain this in full detail with examples. Do not truncate. I want the > comprehensive version. Or for a deep dive on a specific point: > Go deeper on point 2. 200 to 300 words. Include a worked example. You can flip back and forth in the same chat. Brief by default, comprehensive on demand. ## When the rules leak back in ChatGPT sometimes slips back into deference or verbosity when: - You send a long emotional message ("I'm so stuck, nothing is working...") - You correct it twice in a row - The conversation runs past the context limit and earlier instructions get summarized away - You ask an open-ended question that triggers the "be thorough" default If it leaks, re-paste the rules. Twice per long session is usually enough. ## Why "be concise" does not work People type "be concise" and still get a 400-word answer. Three reasons. "Concise" is subjective to the model. Its idea of concise is "I cut the intro." Yours is "one sentence." "Concise" does not override formatting. The model can write a "concise" answer that still has four bullet points and two headers, making it longer than plain prose would be. "Concise" does not block apology phrases. You can get a perfectly concise reply that still opens with "You're absolutely right!" and ends with "Let me know if you need anything else." You need explicit bans on specific phrases, plus explicit format constraints, plus explicit length constraints. All three. Not just "concise." ## A faster way to check Paste a prompt into [FixMyPrompt](https://fixmyprompt.net/try). The rubric flags missing tone constraints, missing length constraints, missing format constraints, model-pleasing phrasing in your input that the model will mirror back, and missing audience anchors. Three free reports per day. No signup. A free account unlocks the rewrite, which adds an explicit tone, length, and format block. ## Related reading - [How to Make AI Writing Sound Less Like AI](/blogs/09-why-ai-sounds-like-ai) - [How to Stop ChatGPT From Making Things Up](/blogs/08-stop-chatgpt-from-making-things-up) - [Your AI Prompt Isn't Working? Here Are the 7 Reasons](/blogs/06-prompt-grader-rubric) - [Why ChatGPT Keeps Misunderstanding Your Instructions](/blogs/07-chatgpt-misunderstanding-instructions) ## Further reading - [Towards Understanding Sycophancy in Language Models](https://arxiv.org/abs/2310.13548) - [Simple Synthetic Data Reduces Sycophancy in Large Language Models](https://arxiv.org/abs/2308.03958) - [Learning to Summarize from Human Feedback](https://arxiv.org/abs/2009.01325) - [System Prompts (Anthropic)](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/system-prompts) - [Be Clear, Direct, and Detailed (Anthropic)](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/be-clear-and-direct) --- ## Gemini Keeps Giving Me Wrong Answers Published: 2026-05-12 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/15-gemini-wrong-answers The same prompt that works on ChatGPT or Claude gets a confused answer from Gemini. Gemini reads prompts differently. Here is what to change. You moved a workflow from ChatGPT to Gemini for the cheaper pricing or the longer context window. The same prompts that produced clean output in ChatGPT now produce wrong facts, off-topic tangents, truncated outputs, ignored format instructions, and extra bullets you did not ask for. You assume Gemini is worse. Sometimes it is. More often, your prompt was secretly tuned to ChatGPT's defaults, and Gemini reads the same text differently. ## Three things Gemini does differently than ChatGPT ### One: Gemini weights the bottom of the prompt heavier ChatGPT pays roughly equal attention to instructions anywhere in the prompt. Gemini's attention skews toward the last instruction it sees. Take this prompt: > Write a 500-word blog post about onboarding. Use bullet points. Include a CTA at the end. Make it formal. Gemini will obsess over "make it formal" because it came last. It may drop "include a CTA" because that instruction was in the middle. Put your most important instructions last. Put style notes first: ``` STYLE: formal, no exclamation points. CONTENT: 500-word blog post about onboarding. FORMAT: bullet points in the body, CTA paragraph at the end. REQUIRED ELEMENTS: (1) intro, (2) 3-bullet body, (3) CTA. All three must be present. ``` The "all three must be present" sentence at the bottom is now what Gemini will optimize against. The output looks much closer to what you wanted. ### Two: Gemini is more literal about format ChatGPT will produce a markdown table when you ask for "a comparison." Gemini gives you whatever it thinks "comparison" means, which is often prose. If you need a specific format, give an example: ``` Output a markdown table with these exact columns: Feature, ChatGPT, Gemini, Winner. Example row: | Pricing | $20/mo | $20/mo | Tie | ``` Gemini follows the example. Drop the example and you get prose. ### Three: Gemini hallucinates differently ChatGPT invents plausible facts. A function that almost exists. A date that is close to right. Gemini hallucinates more on recent events and specific people because its training data weights favor web scraping over targeted RLHF. For anything time-sensitive or person-specific, add this line: > Only state facts you're confident about. If you don't know a specific date, person, or quote, say "I don't know" rather than guessing. In our testing, that single line cuts Gemini hallucination on recent events by around 40%. ## Are you sure you are using the model you think you are? People say "Gemini" like it is one product. It is four very different models. - Gemini Flash: fast, cheap, low accuracy. Good for high-volume routing. Bad for writing. - Gemini 2.5 Flash: speed-optimized. Solid for general tasks. - Gemini 2.5 Pro: the thinking version. Slower, much more accurate. Competitive with Claude Sonnet 4.6 and GPT-5. - Gemini Ultra: top tier. Slow, expensive. Best for complex reasoning. If you are on Flash and the output is weak, "use Pro" is the right answer. If you are already on Pro and the output is weak, the prompt is the right place to look. Check the model selector before blaming the model. ## A prompt structure that fits Gemini ``` [Optional: 1-line role declaration] CONTEXT [everything the model needs. Paste docs, examples, constraints.] STYLE [tone, voice, formatting rules] TASK [what you want done] REQUIRED OUTPUT FORMAT [exact structure with an example] VERIFICATION CHECKLIST - Element 1 must be present - Element 2 must be present - Element 3 must be present - Output must follow the format example exactly ``` The verification checklist at the bottom is what makes Gemini hit your instructions instead of interpreting them creatively. The model now optimizes against the checklist, not against the median of similar prompts. ## When Gemini is actually the wrong tool Some tasks Gemini is weaker at, even on Pro: - Code generation in Python's data-science stack (ChatGPT and Claude lead here) - Long-form fiction with consistent character voice (Claude wins) - Highly conversational tone (ChatGPT is more natural) Where Gemini wins: - Massive context windows. 1M tokens means you can paste a whole codebase. - Image understanding. Strongest of the three for parsing screenshots and diagrams. - Cost. Flash is several times cheaper than equivalent OpenAI or Anthropic models. - Non-English coverage. Pick the model for the task. Do not force every workflow to be model-agnostic. ## A faster way to check Paste a prompt into [FixMyPrompt](https://fixmyprompt.net/try) and select Gemini as the target. The rubric flags important instructions buried mid-prompt (the attention bias), missing format examples, missing "do not guess on facts" lines for time-sensitive tasks, and missing verification checklists. Three free reports per day. No signup. A free account unlocks the rewrite, which reorders your instructions to put critical ones at the bottom. ## Related reading - [Why ChatGPT Keeps Misunderstanding Your Instructions](/blogs/07-chatgpt-misunderstanding-instructions) - [How to Stop ChatGPT From Making Stuff Up](/blogs/08-stop-chatgpt-from-making-things-up) - [Why Claude Refuses Stuff ChatGPT Will Do](/blogs/10-why-does-claude-refuse-things) ## Further reading - [Gemini API Prompting Strategies](https://ai.google.dev/gemini-api/docs/prompting-strategies) - [Gemini Models](https://ai.google.dev/gemini-api/docs/models) - [Long Context with the Gemini API](https://ai.google.dev/gemini-api/docs/long-context) --- ## Why Midjourney Adds Text You Didn't Ask For Published: 2026-05-12 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/16-midjourney-adding-text Midjourney keeps slapping garbled text, fake logos, or random captions onto your images. This is what to add to your prompt to stop it. You generated a logo for a coffee brand. Midjourney returned a beautiful image with the word "CFOFEE" stamped across the cup. You tried again. "COFFEEEE." Then "COFEFE." Then a perfectly drawn but completely fake logo with letters that are not in any alphabet. This is one of the most well-known image-model failure modes. SDXL, DALL-E 3, and Flux all do versions of it. None of them generate legible text reliably, and all of them will add text to your image whether you asked for it or not. Here is what causes it and how to suppress it. ## Why image models invent text Text is just another visual pattern. The model has seen millions of images with text on them. Packaging. Signs. T-shirts. Posters. Book covers. So it learned that coffee cup images often have words on them. When you ask for a coffee cup, the model fills that typical visual region with what looks like text. It has no understanding of letters, spelling, or words. It produces shapes that resemble characters and arranges them where text usually appears. Midjourney v6 and v7 got better at intentional text. You can now ask for `"hello world" written on a sign` and get it mostly right. Unintentional text is still a daily occurrence. ## The fix Add these two things to every prompt where you do not want text: ``` --no text, no letters, no words, no logos, no signs, no labels, no typography ``` And in the prompt body: ``` clean composition with no text or written elements ``` That covers Midjourney (the `--no` flag) and most other image models (the prose constraint). Together they suppress text generation in around 90% of generations. ## When text still appears Some cases leak through even with both fixes. Each has a specific solution. **Fake logo on the bottle.** The model saw "bottle" and pulled a packaged-product pattern. Tell it what the bottle should look like instead: > a plain, unlabeled brown glass coffee bottle on a counter, no markings, no logos, no labels, blank surface **Writing on the building or wall behind the subject.** The model added signage to a streetscape because city backgrounds usually have storefronts. Name a different background: > a person walking on a quiet residential side street, brick wall behind, no shops, no signs, no storefronts **Words floating in mid-air.** Some prompts trigger text more than others. "Abstract" and "concept art" are common culprits. Replace with visual metaphors: > concept art of innovation, visual metaphors only (gears, light, growth shapes), no text, no symbols, no typography **Fake brand text on clothing.** Clothing prompts pull from product photography. Specify plain garments: > a man in a plain solid-color heather-gray hoodie, no logos, no graphics, no brand markings ## When you actually want text If you want intentional text, like the word "OPEN" on a sign: 1. Put the exact text in quotes: `a wooden sign that reads "OPEN" in white painted letters`. 2. Keep it short. One to three words max. Longer text gets garbled. 3. Use Midjourney v6.1 or v7. Earlier versions cannot do legible text reliably. 4. Generate four variations and pick the legible one. Even v7 misses around 30% of the time on text. Anything longer than three words is a bad idea. Generate the image without text, then add text in Figma or Photoshop in post. ## The aspect ratio trick This one surprises people. Certain aspect ratios trigger more text generation than others. `--ar 16:9` (landscape) often adds banner or sign text. `--ar 1:1` (square) tends to add logo-like marks. `--ar 2:3` (portrait) and `--ar 4:5` (social) are usually clean. If you keep getting text leakage and the composition allows it, switch to a portrait ratio. ## A faster way to check Paste an image prompt into [FixMyPrompt](https://fixmyprompt.net/try) with image mode on. The rubric flags missing `--no text` negative prompts, missing "no signs, no logos, no labels" exclusions, subject types known to trigger text leak (bottles, books, signs, clothing, devices), and aspect ratios that raise the risk. Three free image-prompt reports per day. No signup. ## Related reading - [Why Your Midjourney Images Look Generic](/blogs/05-image-prompt-qa-how-to-score) - [How to Get Consistent Characters in Midjourney](/blogs/11-midjourney-consistent-characters) ## Further reading - [Imagen Image Generation Prompt Guide](https://ai.google.dev/gemini-api/docs/imagen-prompt-guide) - [Generating Images with the Gemini API](https://ai.google.dev/gemini-api/docs/image-generation) --- ## Why Does Claude Refuse Stuff ChatGPT Will Do? Published: 2026-05-11 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/10-why-does-claude-refuse-things Claude said no to something ChatGPT happily answered. This is what triggers the refusal, and the four-line preamble that gets the task done. You ask ChatGPT to help you draft a tough customer service reply. Done in 20 seconds. You paste the same prompt into Claude. Now you get this: > I want to make sure I'm approaching this thoughtfully. Could you tell me more about the context? Or worse: > I'm not able to help with that, but here are some general principles around effective communication... The task is fine. The prompt is the problem. ChatGPT and Claude were trained with different defaults for ambiguity. ChatGPT fills in blanks. Claude pauses on them. Same prompt, same task, different reaction. Once you understand that, refusals stop being mysterious. ## The three patterns that trigger refusals Most "Claude refused me" cases fall into one of three shapes. ### Pattern one: it could be read as a real-world harm You typed something benign in your head. Claude is parsing it against the worst plausible interpretation. "Write an angry email about a faulty heater" reads to Claude like it could mean a real interpersonal conflict involving a real person and a real product. The model has no way to tell that you mean a writing sample. Fix the framing. State it is a writing exercise: > I'm a product manager. I need a sample customer complaint email for our QA team's training docs. Write it as the customer would. The product is a faulty space heater. That same email now ships without a hedge. ### Pattern two: impersonation without context "Pretend you're Steve Jobs" feels harmless. Claude sees impersonation, which can be misused for fraud or putting words in a real person's mouth. State the use case and the boundary: > Write in the style of a 2007 Steve Jobs keynote. Punchy, contrarian, big-vision. This is for an internal pitch deck, not attributed to him as a real quote. The output ships. ### Pattern three: dual-use information Some topics help most people and a small number of bad actors. Security research, persuasion techniques, pharmacology. Claude is cautious by default on these. Say who you are and what you'll do with the output: > I'm writing a security awareness email for our IT team. Explain at a high level the most common ways employee passwords get compromised (phishing, reuse, leaked dumps) so I can teach our staff what to watch for. Same information. Now it has a benign frame and a stated audience. ## The four-line preamble that fixes most of it If you only remember one thing, remember this block. Put it at the top of any prompt where Claude has been hedging: ``` Role: I am a [your role]. Context: This is for [specific use, named audience]. Task: [the actual task] Output format: [bullets, draft, table, etc.] ``` A worked example for a website page: ``` Role: I'm a marketing manager at a B2B SaaS company. Context: We're drafting a competitor comparison paragraph for our website. We will not make false claims. Legal will review. The competitor is publicly traded. Task: Write a 100-word paragraph contrasting our async-collaboration features with Slack's, leaning into our offline-first sync. Output format: One paragraph, no headers, neutral professional tone. ``` That goes through Claude 4.5 and 4.6 without a single hedge. ## When Claude is right to refuse Some categories will refuse no matter what context you add, and they should: - Personal data extraction on real people (addresses, phone numbers, identifying details) - Weapons synthesis or operational guides to harm - Sexual content involving minors - Real-time election manipulation If you are working in one of those areas, you do not have a prompt problem. You have a workflow problem. Switch tools. ## Why ChatGPT is not "better" than Claude It is tempting to read all of this as "Claude is more annoying" and move on. The tradeoff is real, though. Claude pauses on ambiguity. That is the same property that makes it stronger on long, careful writing, on code that has to be correct rather than fast, on nuanced editing where ChatGPT will overstep. ChatGPT fills in blanks for you. That is the same property that produces confident-sounding hallucinations and homogenized tone. Once your prompts include the preamble, Claude becomes the easier model for serious work. It just needs the briefing. ## A faster way to check [FixMyPrompt](https://fixmyprompt.net/try) scores prompts against a multi-axis rubric. When you select Claude as the target model, the rubric specifically flags missing role, missing context, missing use case, and impersonation phrasing without disclaimers. Three free reports per day. No signup. A free account unlocks the rewrite, which adds the four-line preamble automatically. ## Related reading - [Why ChatGPT Keeps Misunderstanding Your Instructions](/blogs/07-chatgpt-misunderstanding-instructions) - [Your AI Prompt Isn't Working? Here Are the 7 Reasons](/blogs/06-prompt-grader-rubric) - [Why Just 'Rewriting' Your Prompt Doesn't Actually Fix It](/blogs/04-prompt-optimizer-vs-prompt-qa) ## Further reading - [Claude's Constitution](https://www.anthropic.com/news/claudes-constitution) - [Claude's Character](https://www.anthropic.com/research/claude-character) - [System Prompts](https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/system-prompts) --- ## How to Get Consistent Characters in Midjourney Published: 2026-05-11 Author: FixMyPrompt Team URL: https://fixmyprompt.net/blogs/11-midjourney-consistent-characters The same character looks different every prompt. This is the workflow that actually keeps a character consistent across scenes. You nail the perfect character on the first try. A grizzled space pirate, exactly the energy you wanted. You write a prompt for the next scene. Same pirate, different lighting. Midjourney gives you a totally different person. This is the most common Midjourney frustration. It is also fixable. ## Why the model forgets your character Every Midjourney generation starts from scratch. The model has no memory between prompts. Your character description is a bag of attributes that gets re-rolled each time. "Grizzled, dark hair, scar, leather jacket" matches a million faces. The model picks one. Next prompt, it picks another one. The character was never stored anywhere. It only existed in your head. To get consistency, you have to anchor the new generation to the old one. Three tools do that. The pros use all three together. ## Tool one: `--cref` `--cref` is the real character lock. You point Midjourney at a previous image and say "make the character look like this one." ``` /imagine a grizzled space pirate captain, on the bridge of his ship, holographic star map, cinematic lighting --cref https://i.imgur.com/yourFirstGen.jpg --cw 75 ``` Two flags: - `--cref` takes an image URL. Use a Midjourney-generated image or a clean character reference photo. - `--cw` is character weight, 0 to 100. 100 means strict face lock. Default is 100. If your body keeps looking the same as the reference (same pose, same outfit), drop `--cw` to around 50. That keeps the face but lets the body vary. The most common mistake here is cropping the reference too tight. Midjourney needs to see the full face and shoulders to get a strong lock. A close-up of just the eyes gives bad results. ## Tool two: `--sref` `--sref` locks the visual style. Lighting, color palette, illustration technique. Not the character itself. Use it with `--cref` when you want both face consistency and scene consistency: ``` /imagine a grizzled space pirate captain on the bridge --cref [URL of character] --sref [URL of style reference] --cw 75 --sw 200 ``` `--sw` is style weight, 0 to 1000. The default of 100 works for most cases. Bump to 300 or higher for stricter style matching. ## Tool three: a character sheet block Even with `--cref`, vague descriptions drift. The fewer choices you leave the model, the more consistent the output. Compare these: > Bad: `grizzled space pirate, dark hair, scar` > Good: `weathered 47-year-old human male, deep tan skin, salt-and-pepper short-cropped hair, prominent diagonal scar across left eyebrow, narrow brown eyes, broken nose, neatly trimmed graying goatee, broad shoulders, 6'1", muscular but not bulky, brown leather longcoat with worn brass buttons, three rings on left hand` Every attribute the second version locks down is an attribute the model no longer randomizes. The pros write the character description once, save it as a snippet, and prepend it to every new scene prompt. ## The pro workflow 1. Spend five to ten generations getting your hero character right. Upscale the best one and save the URL. 2. Write the character sheet block once. Save it where you can paste it again. 3. For every new scene: paste the character sheet block, then the scene description, then `--cref --cw 75 --sref