Prompts for cold email work when they make the model write from verified research about one prospect, connect that research to a single problem the prospect’s role plausibly owns, and end with a request small enough to answer in one line. A reliable structure has four parts: research, problem, value bridge and call to action, each handled by its own prompt so you can check the facts before any email is written.
This guide gives each of those prompts in copy-ready form, explains why every instruction is in it, shows an illustrative email built with them, and covers the technical and legal requirements that decide whether a well-written cold email even reaches the inbox.
Why most AI-written cold emails do not get replies
Most AI-written cold emails fail for the same reason most human-written ones do: they are about the sender. A model asked to “write a cold email for our product” will describe the product, list features and ask for a 30-minute call, because that is the most common shape of a sales email in its training data. The recipient sees nothing that shows the sender knows who they are.
The fix is not a cleverer subject line or a friendlier tone. It is better input. Anthropic’s prompting documentation describes the model as responding best to clear, explicit instructions with context, including the reason behind each constraint. For cold email, context means facts about the recipient, and the reason is that relevance is the only thing that earns a reply from a stranger.
There is a second, quieter failure: the model invents personalization. Asked to “reference something about their company”, it may mention a funding round or product launch that never happened. One wrong fact is worse than no personalization, because it proves the email was automated and careless at the same time.
The four-part cold email prompt framework
The framework splits the email into four jobs and gives each job its own prompt. Splitting a task this way is called prompt chaining; the benefit here is that you review the research and the problem statement before they get polished into persuasive prose, which is when errors are hardest to spot.
| Part | Job | What you check before moving on |
|---|---|---|
| 1. Research | Extract facts about the prospect from material you supply | Every fact has a source phrase you can verify |
| 2. Problem | Name one problem their role plausibly owns, linked to a fact | It is framed as a question or observation, not an accusation |
| 3. Value bridge | Connect your offer to that problem as an outcome | No claim you cannot back up |
| 4. Call to action | Ask for one small, easy next step | It can be answered in one line |
Part 1: Research prompt
The research prompt turns material you have read yourself into a short list of usable facts. Paste in the company’s latest news, a job posting, the prospect’s recent post or talk, or the relevant page of their website.
You are a sales researcher preparing one cold email.
Use only the material inside <source>. Do not add anything from memory.
<source>
[Paste: recent company news, a job posting, the prospect's post or
interview, relevant website copy. Include the date of each item.]
</source>
Prospect: [name], [title], [company]
Return:
1. FACTS: up to 5 specific facts, each with the exact supporting phrase
and its date.
2. SIGNAL: the one fact most likely to indicate a current priority for
someone in this role, and why.
3. UNKNOWN: what I cannot conclude from this material.
If nothing in the source is specific or recent enough, say
"No usable signal" and stop.Requiring a supporting phrase for each fact and allowing the model to say “no usable signal” are both techniques from Anthropic’s guide to reducing hallucinations. The dates matter because a “recent” announcement from two years ago makes the email look stale.
Part 2: Problem prompt
The problem prompt links the signal to one challenge that people in the prospect’s role commonly face. One problem, not three; an email that lists several pains reads like a brochure.
Based on the SIGNAL above and the prospect's role ([title]):
Write 3 candidate problem statements, each one sentence. Each must:
- Connect directly to the signal.
- Describe a problem that is common for this role in this situation,
phrased as an observation or a question ("Teams that ... often ...",
"How are you handling ...?").
- Not claim to know their internal situation.
Then recommend one and explain in one sentence why it is the most
relevant to them.The “not claim to know their internal situation” rule exists because statements like “your team is struggling with retention” are presumptuous and usually wrong. A question invites the recipient to tell you, which is the conversation you want.
Part 3: Value bridge prompt
The value bridge is one or two sentences that connect the problem to what you offer, stated as an outcome the reader cares about rather than a list of features.
What we offer: [one sentence]
Proof I can stand behind: [a specific, verifiable result, customer
type, or approach. Leave blank if none.]
Write 2 versions of a 1-2 sentence bridge from the chosen problem to
our offer.
- Describe the outcome for someone in their role, not our features.
- Use only the proof provided. If none was provided, describe the
approach instead of claiming results.
- No superlatives (best, leading, revolutionary, cutting-edge).The “use only the proof provided” line is the most important constraint in the chain. Models readily produce statistics such as “helped clients cut costs by 40%”, and an unsupported performance claim in a commercial email is the kind of statement the FTC expects advertisers to be able to substantiate.
Part 4: Call-to-action prompt and assembly
The final prompt assembles the email and adds a call to action that asks for very little. “Worth a look?” or “Should I send the two-page summary?” can be answered in a word; “Do you have 30 minutes on Tuesday?” asks a stranger to rearrange their week.
Assemble a cold email from the parts above.
Structure:
- Line 1: the signal, stated as a specific observation (no flattery).
- Line 2: the chosen problem.
- Lines 3-4: the value bridge.
- Last line: one low-effort question as the call to action. Offer
something useful (a short summary, an example, a checklist) rather
than a meeting.
Constraints:
- 50-120 words.
- Plain text, no images, no more than one link.
- Write like a person writing to one person: short sentences,
contractions are fine.
- Do not open with "I hope this email finds you well", "My name is",
or "I wanted to reach out".
- Also give me 3 subject lines, under 50 characters, that describe
the email honestly.
Return the email, then a list of every factual claim it makes and
where each one came from.The closing request for a claims list is your final check. Any claim without a source goes before you send. For subject lines specifically, there is a dedicated set of prompts for sales email subject lines.
Worked example: one email built with the framework
Everything in this example is invented for illustration. The sender runs a small consultancy that sets up on-call processes and incident runbooks for engineering teams. The prospect, “Marco”, is Head of Engineering at a fictional logistics software company whose careers page shows three open site reliability roles, and whose engineering blog recently described a two-hour outage and the lessons learned.
Before, from a one-line prompt (“write a cold email to a Head of Engineering about our incident response consulting”):
Subject: Transform Your Incident Response!
Hi Marco,
I hope this email finds you well. My name is Alex and I'm the founder of
a leading incident response consultancy. We help innovative companies
like yours reduce downtime by up to 60% with our cutting-edge
methodology. Do you have 30 minutes next week to discuss how we can
help your team?After, from the four prompts with the blog post and job listings as source:
Subject: Your outage write-up and the 3 SRE openings
Hi Marco,
Read your team's post on the routing outage. The point about the
runbook pointing to a dashboard that had been renamed stuck with me.
With three SRE roles open, I'd guess a lot of on-call knowledge is
about to move into new heads. How are you planning to get new hires
on the rotation without relearning that kind of lesson?
I help engineering teams turn post-incident notes into runbooks that
new on-call engineers can follow on day one.
Want me to send a one-page template we use for that handover?
AlexThe before version contains an unsupported statistic (“up to 60%”), two empty superlatives and a large ask. The after version is built entirely on two facts the sender read, raises one problem as a question, and asks for a one-word reply. It is also honest about what is a guess (“I’d guess”).
Follow-ups that add new information
A follow-up should exist only if it adds something the first email did not. “Just bumping this to the top of your inbox” gives no new reason to reply.
The prospect did not reply to: "[paste first email]".
Write follow-up 1 (to send about a week later) and follow-up 2 (a
polite close).
Follow-up 1:
- One new element only: a short example, a relevant public resource,
a different angle on the same problem, or a narrower question.
- Under 70 words. Reply in the same thread.
Follow-up 2:
- Under 50 words. Say this is the last note, leave one easy way to
pick it up later.
Neither may repeat the first email's wording or guilt-trip the reader.For a fuller treatment of sequences and timing, see the guide to sales follow-up emails.
Adjusting tone for the recipient’s role
The same offer should be described differently to different roles, because each role measures success differently. Rather than asking the model to “change the tone”, tell it what the reader is accountable for.
Rewrite this email for a [new title]. This person is measured on
[e.g. pipeline and campaign results / budget and risk / delivery
speed and uptime]. Keep every fact. Change the problem framing and the
outcome language to match what they are measured on. Keep the length.Describing what the reader is accountable for gives the model something concrete to adapt to, which works better than adjectives like “creative” or “executive-friendly”.
What has to be true before you hit send
A well-written cold email still fails if it lands in spam or breaks the law. Two sets of rules matter most for senders emailing US recipients and Gmail users.
CAN-SPAM applies to B2B email
The US CAN-SPAM Act covers commercial email, and the FTC’s compliance guide states that it has no exception for business-to-business messages. In summary, the guide requires accurate header information, a subject line that reflects the content, identification of the message as an ad, a valid physical postal address, a clear way to opt out, and opt-out requests honored within 10 business days. Other countries have their own rules, some stricter, so check the law where your recipients are. For a detailed walkthrough, see what an AI-drafted marketing email still needs to pass CAN-SPAM.
Gmail’s sender requirements
Google’s email sender guidelines require all senders to Gmail accounts to authenticate with SPF or DKIM, have valid forward and reverse DNS, use TLS, and keep the spam rate reported in Postmaster Tools below 0.3%. Bulk senders (5,000 or more messages a day to Gmail) also need SPF, DKIM and DMARC and one-click unsubscribe for marketing messages. Cold outreach from a domain without authentication, or sent in volume from a brand-new domain, is the most common reason good copy never gets seen; the article on technical mistakes that get AI-personalized cold emails flagged as spam covers the details.
A final pre-send check
- Could this email be sent to anyone else with only the name changed? If yes, go back to part 1.
- Is every fact about the prospect checked against something you read yourself?
- Is every claim about your results something you can document?
- Is there exactly one ask, and is it small?
- Are the opt-out and postal address in place, and is your sending domain authenticated?
Questions about AI and cold email
Should I let an AI tool send cold emails automatically?
Drafting at scale is reasonable; sending without review is risky. The research step only works if a person checks the facts, and automated high-volume sending is what pushes spam rates up. A sensible split is AI for research and drafts, a person for the final read and the send.
What if I cannot find anything specific about a prospect?
Then that prospect may not be ready for a personal email. You can still write to them using a role-level observation, but be honest that it is general. Forcing a fake personal detail is worse than a well-targeted general email.
