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:
- Are the most important rules in the last three lines? Move them if not.
- Did you say what to do, not what not to do? Flip the negatives.
- Are there contradictions? Pick one side.
- Did you specify the output format? Length, structure, example.
- 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. It scores all five of the above and tells you which one is bleeding the most signal.
Three free reports per day. No signup.