You typed "write me a cold email to sell my SaaS to CTOs." ChatGPT produced this:
Subject: Streamline Your Operations Today!
Hi [Name],
I hope this message finds you well. I wanted to reach out to introduce our innovative platform that helps businesses like yours streamline operations, reduce costs, and unlock new growth opportunities...
That email has a 0.4% reply rate. We know because every cold email tool reports their version of that template and the data lines up.
ChatGPT defaulted to the worst possible format because 80% of the cold emails in its training data look like that. The model knows how to write better. It just needs to be told to.
What kills the default version
Three things, every time.
The subject line is generic. "Streamline Your Operations Today!" reads identical to fifty thousand emails the recipient deleted last month.
The opener is "I hope this message finds you well." This is the single most spam-flagged opening line in the inbox. Every spam filter is tuned for it. Every human reader has trained themselves to skim past it.
The body is a feature dump. Three sentences of buzzwords with no specific reason this recipient should care.
A real reply-getting cold email looks the opposite. Specific subject. Specific opener tied to the recipient. One sharp pain. One yes/no question.
The prompt that actually works
Copy-paste this into ChatGPT or Claude. Fill in the bracketed bits with your real info:
You are writing a cold email. Follow these rules with no exceptions.
CONTEXT
- Sender: [your name + 1-line title]
- Sender's product: [1 sentence on what you sell, no features]
- Recipient: [name, title, company]
- Specific recipient observation: [something true about them. A
recent post, hire, funding round, or public initiative. Not
generic. Not in 50,000 other cold emails.]
RULES
- Subject line: 5 words max. Lowercase. Specific. Never use
"quick question", "[their company]", "synergy", "opportunity",
"growth", or "streamline".
- Opening line: do not say "I hope this finds you well" or any
variation. Open with the recipient-specific observation. Make
it sound like you actually read about them.
- Body: 3 sentences max.
- Sentence 1: the observation, paraphrased.
- Sentence 2: a sharp observed pain. Something most companies
at their stage struggle with. Be specific. Not "growth
challenges."
- Sentence 3: one sentence on what you do, framed as a possible
fix for that pain. Plain words. No buzzwords.
- CTA: one sentence asking a yes/no question they can reply to
in five seconds. Not "do you have 15 minutes" or "open to a
quick call".
- Signature: name only. No title. No company. No logo.
- Total length: under 70 words.
OUTPUT FORMAT
Subject line on first line, then body. Nothing else. No explanation.
Paste in real context (recipient and observation) and ChatGPT produces something close to this:
Subject: noticed the kubernetes migration
Saw your engineering blog post about the recent k8s migration. Congrats on getting off ECS.
Most teams hit a wall around month three when prod traffic exposes gaps in dev-cluster observability, and on-call burns out.
We instrument k8s pods at the system-call level so engineers know which pod is causing pager noise without writing custom Prometheus exporters. Worth a two-minute reply on whether on-call is currently painful for you?
Alex
In our internal testing, that style of email lands around 12% reply rate. The difference is not the model. The difference is the rule list.
The five rules that do most of the work
If you remember nothing else from the prompt above:
Keep the subject under five words and lowercase. It will look like a human typing, not a marketing template.
Ban "I hope this finds you well." Replace it with a recipient-specific observation. This single change cuts spam-folder rates significantly.
One sharp observed pain, not three vague benefits. Specific beats comprehensive.
The CTA asks for a reply, not a meeting. "Worth a two-minute reply?" beats "Do you have 15 minutes?" every time.
Total length under 70 words. Anything over 100 dies in the preview pane.
Why this works on a model that writes the bad version by default
Buried in ChatGPT's training data are plenty of well-written cold emails. They are just outnumbered. Without rules, the model statistically pulls the average pattern. With explicit format constraints, it pulls a much better one. Same model. Different prompt.
What a perfect email cannot fix
A great prompt cannot save you from a bad list. Cold outreach to people who do not buy your product, no domain warm-up, fake observations like "I see you work in tech." None of that is a prompt problem. Fix those separately.
A faster way to check
Paste your cold email prompt into FixMyPrompt. The rubric flags missing recipient observations, dead-on-arrival openers, vague pain claims, and meeting-request CTAs. The rewrite enforces the five rules above.
Three free reports per day. No signup.