Short answer: The best AI prompt for cold email doesn’t ask the AI to write an email — it first makes the AI
research the prospect, then write a first line that proves you did that research. Personalization is what separates a 15% reply rate from a 1% one. Run a research prompt first, then feed those findings into the writing prompt. Output quality is capped by input specificity.
TL;DR — Key Takeaways #
Never ask AI to write first. Run Prompt 1 (research) before Prompt 2 (writing). The research output becomes the input for everything else.The average B2B cold email reply rate is ~3.4%. Signal-based personalized campaigns hit 15–25% — roughly 5x the average.Small batches beat blasts. Campaigns under 50 recipients average 5.8% replies vs. 2.1% for large blasts.Follow-ups generate the meetings. Most replies arrive on emails 2–4, not the first send.Deliverability is the floor. Without SPF, DKIM, and DMARC, the best email still lands in spam.
✔ Best for Founders, SDRs, AEs, and agency owners sending 10–500 cold emails weekly who want to move beyond template spray-and-pray.
✕ Skip if You’re sending inbound follow-ups to warm leads (different playbook), or looking for software recommendations rather than prompt systems.
On this page #
Why do most AI-generated cold emails fail? #
Because the prompt contains no information the AI couldn’t have guessed. If your prompt is “write a cold email to a SaaS CEO about our analytics tool,” the model has nothing specific to work with. It produces the statistical average of every cold email in its training data — which is exactly the email your prospect deletes.
AI doesn’t fix a weak message. It scales whatever you give it. Feed it generic input and you get generic output, faster. The entire skill is front- real, specific, verifiable detail about the prospect before you ask for a single word of copy.
What actually makes a cold email get a reply? #
Four things, in this order of impact:
| Factor | Why it matters | Example |
|---|---|---|
| Relevance to a current situation | ||
| Something happening at their company right now — a hire, a launch, a job posting, a public complaint. | “Four support-rep openings in a month usually means something broke.” | |
| Specific, small ask | ||
| A low-friction commitment that feels doable. “Worth 15 minutes Thursday?” beats “let’s discuss synergies.” | “Worth 15 minutes Thursday to see if the same approach fits?” | |
| Brevity | ||
| Under 100 words total. Readable in a phone notification preview without expanding. | 61-word example below | |
| A reason it’s them | ||
| Not their job title or industry — a specific signal about their company this month. | Their Trustpilot reviews, job postings, public statement, LinkedIn post. |
Notice that none of these are “clever subject line” or “compelling value prop.” Those matter, but they’re multipliers on a message that’s already relevant. A brilliantly written email about something the prospect doesn’t care about is still deleted.
The 4-prompt cold email system #
The system runs in sequence. Each prompt’s output becomes the next prompt’s input. Don’t skip step one.
Prompt 1 — The personalization research prompt
This is the highest-leverage prompt in the entire guide. It does not write anything. It finds you the raw material.
You are a B2B sales researcher. I am about to send a cold email to
[PROSPECT NAME], who is [JOB TITLE] at [COMPANY NAME]
(website: [COMPANY URL]).
I sell [YOUR PRODUCT/SERVICE] which helps [TARGET CUSTOMER TYPE]
achieve [SPECIFIC OUTCOME].
Research this person and company, then give me:
1. THREE specific, recent, verifiable facts about the company
(last 6 months): funding, hires, launches, expansions,
partnerships, press, or public strategy shifts.
2. TWO signals that suggest they may have the problem I solve.
Cite the evidence for each.
3. ONE thing about this specific person: recent post, podcast,
talk, interview, article, or career move.
4. The single most likely reason they would ignore my email.
Rules:
- Only include facts you can point to a source for.
- If you cannot verify something, write "NOT FOUND."
- Rank the facts by how relevant each is to the problem I solve.
Why it works: it forces sourcing, bans invention (the number-one failure mode), and asks for the objection up front so you can pre-handle it. The “NOT FOUND” instruction is critical — without it, models fill gaps with plausible fiction.
Verify before you send. Treat this prompt’s output as leads to check, not facts to paste. One hallucinated detail (“congrats on the Series B”) destroys the credibility the entire email depends on. Use a model with live web access and click the sources.
Prompt 2 — The first-line prompt
The first line is the whole email. If it doesn’t earn the second line, nothing else in the message exists.
Write 5 different opening lines for a cold email to [PROSPECT NAME],
[JOB TITLE] at [COMPANY].
Use only these researched facts:
[PASTE THE OUTPUT FROM PROMPT 1]
Requirements for each line:
- Maximum 15 words.
- Reference one specific researched fact — not their job title,
not their industry.
- No compliments. No "I was impressed by." No "I came across
your profile."
- No mention of me or my company yet.
- Written like a human peer, not a vendor. Plain words.
- Each of the 5 should use a DIFFERENT fact or angle.
After each line, add a one-sentence note on why a busy
[JOB TITLE] would keep reading.
Why it works: the banned-phrases list kills the tells that mark an email as automated. Forcing five variations gives you a real choice instead of accepting the first instinct.
Prompt 3 — The full email prompt
Write a cold email using this structure:
OPENING: [PASTE YOUR CHOSEN LINE FROM PROMPT 2]
Then:
- OBSERVATION: connect that fact to a problem companies like
theirs typically face. One sentence. No stats I can't verify.
- CREDIBILITY: one sentence on how we helped
[SIMILAR COMPANY] achieve [SPECIFIC RESULT WITH A NUMBER].
- ASK: a single, low-commitment question. Ask for a reply or
15 minutes — never both.
Hard constraints:
- Under 90 words total.
- Fifth-grade reading level.
- No adjectives like "innovative," "cutting-edge," "seamless,"
"revolutionary," "game-changing."
- No "I hope this email finds you well," "quick question,"
"just following up," "circling back," "touching base."
- No bullet points. No bold. Plain text only.
- Do not use my company name more than once.
- Subject line: under 5 words, lowercase. Give me 3 options.
Context:
- I sell: [YOUR PRODUCT/SERVICE]
- To: [TARGET CUSTOMER]
- Outcome we deliver: [SPECIFIC MEASURABLE OUTCOME]
- Proof point: [CASE STUDY / METRIC / CLIENT NAME]
- Their likely objection: [FROM PROMPT 1, ITEM 4]
Why it works: the banned-word list does more work than the instructions. Those phrases are the fingerprints of mass outreach. Removing them alone makes an email read as human. The word cap forces the model to cut the throat-clearing it defaults to.
Prompt 4 — The follow-up sequence prompt
Most meetings come from emails 2 through 4. This is the prompt most people never run.
Write a 4-email follow-up sequence for the cold email below.
The prospect has not replied.
[PASTE YOUR EMAIL FROM PROMPT 3]
Sequence rules:
- Email 2 (day 3): NEW information. A different angle on the
same problem. Do not reference the first email.
- Email 3 (day 7): social proof. A one-line customer story
with a number, from a company like theirs.
- Email 4 (day 12): a genuinely useful resource with no ask
attached. Give something away.
- Email 5 (day 18): the break-up. Short, warm, no guilt.
Explicitly close the loop.
Constraints for all:
- Under 60 words each.
- Never say "just following up," "bumping this," "did you see
my last email," or "in case you missed it."
- Each email must stand alone and deliver value even if they
never read the previous ones.
- Vary the sentence structure so the sequence doesn't read
like one template.
Why it works: most follow-ups are nags — they add pressure without adding information. This sequence adds a new reason to reply each time, which is why it converts.
Good prompt vs. bad prompt: the difference in practice #
| Weak prompt | Strong prompt | |
|---|---|---|
| Input | ||
| “Write a cold email to a SaaS CEO about our analytics tool” | Research output + named company + verified signal + proof point + word cap | |
| Personalization | ||
Merge-tag level: [First Name] , industry |
||
| Situation level: their hire, launch, or job posting | ||
| Constraints | ||
| None | Word count, reading level, banned phrases, one ask | |
| Ask | ||
| “Let me know if you’d like to chat!” | “Worth 15 minutes Thursday?” | |
| Output | ||
| The same email 10,000 people sent | An email only you could have sent | |
| Realistic reply rate | ||
| Around 1% or below | 8–15%+ on a clean, well-targeted list |
Reply-rate ranges reflect published 2026 benchmarks, where the B2B average is ~3.4%, top performers reach 8–12%, and signal-based personalized campaigns reach 15–25%.
A real example: input → output #
Scenario: You sell a customer-support automation tool. Your prospect is the VP of Customer Experience at a mid-size ecommerce brand.
Prompt 1 output (research):
- Company posted 4 support-rep job openings in the last 30 days (careers page)
- Launched into the UK market in May (press release)
- Trustpilot reviews from the last 60 days repeatedly mention slow email response times
- VP posted on LinkedIn about “scaling support without scaling headcount”
- Most likely reason they’d ignore this: they’re mid-vendor-evaluation and drowning in pitches
Prompt 2 output (chosen first line):
“Four support rep openings in a month usually means something broke.”
Prompt 3 output (full email):
Subject: four openings
Four support rep openings in a month usually means something broke.
Usually it’s volume growing faster than the team — and after a UK launch, that tracks. The reviews mentioning slow replies suggest the same thing.
We helped Brightline cut first-response time from 14 hours to under 1 without adding headcount.
Worth 15 minutes Thursday to see if the same approach fits?
Every sentence is grounded in something verifiable. The prospect can tell in four seconds that this wasn’t blasted to 5,000 people — and that’s the only thing that earns a reply.
Level-up: the objection-preemption prompt #
This is the prompt almost nobody runs, and it’s where the real gains are. Elite outreach anticipates the specific reason this prospect will say no and defuses it before it forms.
Act as [PROSPECT NAME], [JOB TITLE] at [COMPANY]. You receive
20+ vendor emails a day and delete most in under 2 seconds.
Here is an email you just received:
[PASTE YOUR EMAIL]
Respond in three parts:
1. YOUR HONEST REACTION — the actual thought in your head as
you read the first line. Be blunt. If you'd delete it, say so
and say exactly which word or phrase triggered that.
2. THE THREE OBJECTIONS you'd raise if you did reply. Rank by
likelihood.
3. THE REWRITE — rewrite the email so that it survives your own
objections. Change only what's necessary. Keep it under
90 words.
Then step out of character and tell me: what single change
would most increase the odds of a reply?
Why this unlocks wins: role-play inverts the model’s default bias. Asked to write an email, an AI optimizes to please you. Asked to receive one as a jaded buyer, it optimizes to critique — and it becomes remarkably good at spotting the exact phrase that reads as salesy.
Stacking tip: run this twice. Once as the prospect, once as “a skeptical CFO who has to approve any spend.” Different personas surface different objections.
Before you send: the deliverability floor #
The best email in the world doesn’t matter if it lands in spam. These are prerequisites, not optimizations.
Authenticate your domain. SPF, DKIM, and DMARC are required by Google, Yahoo, Microsoft, and Apple. Compliant senders average ~89% inbox placement; non-compliant senders see 22–34% routed to spam.Keep spam complaints under 0.1%. The enforced ceiling is 0.3%; past that, deliverability degrades and recovers slowly.Verify your list. Verified lists get roughly 2x the reply rate of unverified ones. Bounces need to stay under 2%.Send small batches. Under 50 recipients per campaign averages 5.8% replies vs. 2.1% for large blasts. This is also where real personalization becomes feasible.Use a separate sending domain. Never burn your primary domain’s reputation on cold outreach.Know the law where you’re sending. CAN-SPAM (US) requires accurate headers and a working opt-out. GDPR (EU/UK) requires a lawful basis. Rules vary by country.
Which AI model should you use? #
| Task | Best fit | Why |
|---|---|---|
| Prospect research(Prompt 1) | ||
| Any model with live web search — ChatGPT with search, Perplexity, Claude with web access | ||
| Research is worthless without current data. Training data alone will invent facts. | ||
| First lines & copy(Prompts 2–3) | ||
| Claude or ChatGPT | Both handle tight constraints and banned-word lists well. Claude defaults to less filler. | |
| Role-play critique(Level-up) | ||
| Claude or ChatGPT | Both sustain an adversarial persona without softening into flattery. | |
| Bulk variations at scale | ||
| Whichever you can call via API | Cost and throughput matter more than nuance once templates are locked. |
All prompts are model-agnostic — they work in ChatGPT, Claude, and Gemini. The only hard requirement: Prompt 1 must run on a model with live web access.
Frequently asked questions #
Can ChatGPT write cold emails that actually get replies?
Yes, but only when the prompt contains researched, specific information about the prospect. ChatGPT from a generic prompt performs at or below the ~3.4% B2B average. The same model, given a research brief with verified company signals, produces emails that perform at 15–25% reply rates for signal-based outreach.
What’s a good reply rate for cold email in 2026?
The B2B average is ~3.4%, with a ~27.7% open rate. Normal: 1–5%, strong: 8–12%, elite signal-based campaigns: 15–25%. Reply rate is more reliable than open rate, which is inflated by Apple Mail Privacy Protection.
How long should a cold email be?
Under 90 words. It should be fully readable in a phone notification preview without expanding. Follow-ups should be under 60 words. Length is the most common self-inflicted wound — every sentence past the ask reduces reply odds.
How many follow-ups should I send?
Four after the initial email, spaced at days 3, 7, 12, and 18, with the last being a break-up email. Most replies arrive on emails 2–4, so stopping after one send wastes most of your list’s value.
Will prospects be able to tell my email was written by AI?
They can tell when it’s unedited AI. The tells are predictable: “I hope this email finds you well,” “I came across your profile,” adjectives like innovative and seamless, bullet points, and throat-clearing. Ban those in your prompt and always edit the output. The goal is AI-assisted, not AI-authored.
Is cold email still legal?
In most jurisdictions, yes, with conditions. US CAN-SPAM requires accurate headers, a physical address, and a working opt-out. GDPR (EU/UK) requires a lawful basis for processing. Rules vary by country — verify yours. This is general information, not legal advice.
What’s the single biggest mistake in AI cold email?
Letting the model invent facts. A hallucinated detail — congratulating someone on a funding round that never happened — is worse than no personalization because it proves the email was automated and careless. Instruct the model to write “NOT FOUND” rather than guess, and verify every claim before sending.
Should I personalize every email or use templates?
Both. Use a locked template structure and personalize only the first line and observation sentence — roughly 25 words per email. That’s achievable at 50 emails/day and delivers most of the lift. Full bespoke emails don’t scale; zero personalization doesn’t work.
Get the Cold Email Prompt Pack #
Download the 10-Email Swipe File + Prompts
All four system prompts, the objection-preemption prompt, the banned-phrase checklist, and a 10-email swipe file ready to customize. Copy-paste ready.
Written by the Narracomm team
Narracomm is a communications and content strategy team that helps business owners, operators, and founders use AI to produce clear, credible, high-performing content. We build and test prompt systems inside real client work — across sales, marketing, and operations — and update them as models change. [Add specific credentials, years of outbound experience, and a named reviewer here to strengthen E-E-A-T.]
Sources & further reading #
Instantly — Cold Email Benchmark Report 2026Amplemarket — 2026 cold email benchmarksCleanlist — Cold Email Response Rates 2026PowerDMARC — Bulk Email Sender Requirements (2026)PowerDMARC — Google & Yahoo Authentication RequirementsRed Sift — 2026 Bulk Email Sender Requirements Checklist
Last reviewed and updated: July 25, 2026 · Model versions verified against current releases. Next review due within 14 days.