AI Cold Email Prompts: A 2026 Prompt Library

Bad prompts produce ChatGPT-flavored slop. You know the kind: "I hope this email finds you well," followed by three paragraphs of vague value props that could apply to any business on earth. Good prompts produce cold emails that prospects actually reply to — because they sound human, they reference something real, and they ask for something small. The difference is not the model. It is the prompt.

This post is a 2026 prompt library for ai cold email prompts. Every prompt below has been used at scale, refined against real reply data, and pressure-tested across industries. Copy them, adapt them, and feed them structured prospect data — ideally exported from MapsLeads — to get outputs that do not read like a robot wrote them in a hurry.

The anatomy of a prompt that works

Before we get to the templates, internalize the structure. Every prompt in this library follows the same five-part skeleton:

1. Role — who the AI is pretending to be (a senior SDR, a copywriter, a classifier).
2. Context — the data and situation. This is where MapsLeads fields plug in.
3. Constraints — word limits, tone rules, things to avoid.
4. Few-shot examples — one or two reference outputs that anchor the style.
5. Format — exactly what the model should return. JSON, plain text, three paragraphs.

If a prompt is missing any of these five, the output will drift. The most common failure is missing constraints — the model defaults to its training distribution, which is enthusiastic LinkedIn-speak. Constraints save you.

1. The research prompt

Before writing anything, the AI needs to know the prospect. This prompt summarizes a business into the three or four facts that can carry an opener.

You are a B2B research analyst. I will give you structured data about a local business. Return exactly four bullets: (1) what they do in plain English, (2) one signal they care about reputation, (3) one signal they are growing or struggling, (4) one specific public anchor — a recent review keyword, a photo type, or a service category — that a stranger could plausibly reference. No adjectives like "thriving" or "innovative". 60 words total maximum.

Data: name, category, rating, review_count, recent_review_keywords, photos_count, website, phone.

Example output for a dental clinic: "Family dental clinic in Austin. Reputation matters: 412 reviews, 4.8 rating. Growing: review velocity tripled in the last 90 days. Anchor: recent reviewers repeatedly mention 'Dr. Patel's chairside manner with kids' — a specific human, not a generic compliment."

That last bullet is the gold. It is what your opener will reference.

2. The opener prompt

One line. One specific anchor. No flattery.

You are writing the first line of a cold email. Use the anchor I provide and write exactly one sentence, 18 words or fewer, that proves you actually read about this business. Forbidden phrases: "I came across", "I hope this finds you well", "impressive", "love what you're doing", "stood out". Use a casual, observational tone.

Anchor: [recent_review_keywords], [category], [rating].

Example output: "Saw the run of reviews calling out Dr. Patel for being good with anxious kids — that is a hard thing to train for."

3. The body prompt (PAS framework)

Problem, Agitation, Solution. Three short paragraphs. No more.

You are a senior B2B copywriter. Write the body of a cold email using the PAS framework. Paragraph 1: name a specific operational problem this business likely has, given their [category] and [review_count]. Maximum 25 words. Paragraph 2: agitate by naming the cost of that problem in concrete terms — hours, dollars, missed bookings. Maximum 25 words. Paragraph 3: introduce [our_product] as the fix, in one sentence, with one number. Maximum 25 words. No bullet points. No bold. Conversational tone.

Example output: "Clinics your size usually field 30 to 50 new-patient calls a week, and most never get logged. Every unlogged call is a no-show waiting to happen — we estimate the average single-location practice loses about $4,200 a month to that gap. Caller-ID drops every inbound into your CRM with the review history attached, so your front desk knows who is calling before they pick up."

4. The CTA prompt

There are two CTAs that work in 2026: the interest CTA and the ask CTA. Prompt for both, A/B test them.

You are closing a cold email. Generate two CTAs. (A) Interest CTA: a single question that requires only "yes" or "no" to answer, asking if the topic is worth a conversation. (B) Ask CTA: a specific 15-minute meeting request with two suggested time windows. Both must be under 20 words. No "circle back". No "touch base".

Example A: "Worth a five-minute look at what your last 100 missed calls were worth?" Example B: "Got 15 minutes Thursday at 10:30 or Friday at 2 to walk through your numbers?"

5. The reply classifier prompt

Once replies start coming in, you need to triage them at machine speed. This prompt routes every reply into one of four buckets.

You are an email classifier. Read the reply and return JSON with two keys: "label" (one of: positive, objection, ooo, unsubscribe) and "confidence" (0 to 1). Definitions. Positive: any signal the prospect wants to continue, including questions, meeting acceptance, or "send more info". Objection: pushback, "not now", "not a fit", "wrong person" with no warm handoff. OOO: out-of-office auto-replies and vacation messages. Unsubscribe: any request to stop, including "remove me", "take me off", or hostile language.

Example input: "Not now, but ping me in Q3." Example output: {"label":"objection","confidence":0.86}

This single prompt eliminates the need for 80% of human inbox triage.

6. The negative example prompt

Sometimes the most useful prompt is one that defines what NOT to write. Use it as a quality gate.

You are a critic. I will paste a cold email. Return a list of every cliche, AI-tell, or generic phrase. Specifically flag: "I hope this email finds you well", "I came across your", "wanted to reach out", "circle back", "synergy", "leverage", "unlock", "game-changer", any sentence that could apply to any business in the world, any compliment without a specific reference, and any paragraph longer than 30 words. If the email is clean, say "clean".

Run every generated email through this gate before it sends. Reject anything that comes back with more than one flag.

7. The personalization prompt with MapsLeads fields

This is where structured data turns AI from a hallucination machine into a sniper rifle.

You are personalizing a cold email. The prospect data has these exact fields: business_name, category, rating, review_count, recent_review_keywords (array of 3 to 5 phrases), photos_count, website, phone, contact_first_name. Replace every [brace] placeholder in the template below with real values. If recent_review_keywords contains a person's name (e.g. "Dr. Patel"), use it. If photos_count > 100, mention they are visually invested. If rating < 4.2, soft-pedal — do not lead with reputation.

Few-shot anchors are critical here. Give the model two prior examples of correctly personalized emails and one example of a wrongly personalized one (e.g., where the AI invented a fact). The contrast trains the model to stay grounded.

8. The subject line prompt

Subject lines deserve their own prompt because the rules are different. Short, lowercase, curiosity, no clickbait.

Generate 5 subject lines for the email below. Rules: 4 to 7 words each, lowercase, no punctuation except a question mark or comma, no emojis, no "quick question", no "[first_name]" if it makes the line awkward, must hint at the email's content without giving the answer.

Example outputs: "the dr patel reviews thing"; "missed calls at your clinic"; "what 412 reviews are not telling you"; "a small dental ops question"; "noticed something about your reviews"

9. The multi-language prompt

In 2026 the cleanest way to localize is not "translate this email" — it is "rewrite this email as a native speaker would have written it from scratch."

You are a native [target_language] B2B copywriter. I will give you a cold email in English. Do not translate. Rewrite it from scratch as a native [target_language] speaker would write to a [country] prospect, preserving the structure (opener, PAS body, CTA), the specific facts, and the word counts. Adapt idioms, formality level, and date formats to the target locale.

This produces output that does not read as translated. For German B2B, it knows to use "Sie". For Brazilian Portuguese, it knows the casual register. For Japanese, it knows to soften the CTA.

10. The tone-of-voice transfer prompt

If your founder writes in a distinctive voice, capture it.

You are imitating a specific voice. Below are 5 emails the founder has written. Study the sentence rhythm, vocabulary, punctuation habits (Oxford comma yes/no, dashes vs parentheses), and recurring tics. Now rewrite the draft email in that exact voice, preserving meaning but matching cadence. Do not invent new claims.

Paste five real emails. The output will sound like the founder, not like GPT.

How MapsLeads fields plug into the prompts

Every prompt above assumes you have clean, structured prospect data. This is where MapsLeads becomes the upstream half of your AI cold email stack. When you export from MapsLeads, you get columns named exactly the way the prompts above expect them: rating, review_count, recent_review_keywords, photos_count, category, contact_first_name. Your AI agent reads those columns directly — no scraping, no guessing, no hallucinated facts.

The field that does the most work is recent_review_keywords. It is the difference between "I see you have great reviews" (slop) and "saw the run of reviews calling out Dr. Patel for being good with anxious kids" (specific). Without that field, the AI has nothing real to anchor to.

Photos_count is the next-best signal. A business with 300+ photos is visually invested and probably proud of their physical space — that is an opener. A business with 4 photos has a different profile, and the AI should not pretend otherwise.

The credits work like this: 1 credit Base for the business record, +1 Contact Pro for verified email and direct contact name, +1 Reputation for review_count, rating, and recent_review_keywords, +2 Photos for the photo intelligence. The full enrichment runs you 5 credits per prospect — well under what a single human-researched lead costs anywhere else.

The workflow is simple: Search in MapsLeads, toggle Reputation (+1 credit), export the CSV, hand it to your AI agent, and the agent reads the structured columns straight into the prompts above. No middleware. No prompt engineering on raw HTML.

For the full picture of how this fits into an automated outbound stack, read our AI SDR complete guide 2026 and cold email personalization at scale.

Common mistakes

The five mistakes we see weekly. First: prompts with no constraints. The model defaults to enthusiasm. Second: prompts that ask for paragraphs instead of word counts. "Short" means nothing — "25 words maximum" means something. Third: prompts that hand the AI a website URL and expect it to "research." Models are not browsers; feed them structured data. Fourth: skipping the negative-example gate. Every email should pass the critic prompt before it sends. Fifth: using the same prompt for every industry. Dental clinics and SaaS founders deserve different angles, even if the structure is identical.

FAQ

How do I prompt AI to write cold emails? Use the five-part skeleton: role, context, constraints, few-shot examples, format. Skip any of those and the output drifts.

What is the best AI cold email prompt? There is no single best prompt — there is a chain. Research prompt feeds the opener prompt, which feeds the body prompt, which feeds the CTA prompt. The classifier handles replies. Each does one thing well.

Should I use ChatGPT or Claude for cold emails? Both work. Claude tends to follow constraints more literally and produces fewer cliches out of the box. GPT is more creative but needs a tighter critic loop. Test both with the same prompts on the same data.

How do I avoid AI-flavored email? Forbid the cliches by name in your prompt, cap every paragraph at 25 words, and run every output through the negative-example critic prompt before sending. If you cannot point to a specific anchor in the email, the AI is guessing — kill the line.

Do I need few-shot examples? Yes. Two well-chosen examples beat any amount of instruction text. Use real emails that got real replies.

What about subject line testing? Generate 5 with the subject prompt, send each to a 50-prospect cohort, keep whatever beats baseline open rate by 3 points or more.

Start sending emails that do not sound like AI

Good prompts plus structured data plus a critic loop. That is the entire trick. The prompts in this library are the starting point — adapt them, A/B test them, and feed them real prospect data.

If you need ready-made templates to plug these prompts into, see our cold email templates for B2B SaaS. To pull the structured prospect data the prompts depend on, check pricing or get started free with 50 credits. That is enough to enrich 10 fully-loaded prospects and run your first AI-written sequence end to end.