AI cold email prompting works best when the human supplies real observations, experiences, and context. Give those to the model and ask it to place them in a structure informed by account research. The human provides the substance, and AI handles the expression.1
A prompt that only asks for a personalized email leaves too many decisions open. Give the model a clear job, a bounded input, a defined sender, a specific goal, and rules for what the output may claim. That structure gives you something to inspect before any message reaches a prospect.
Start with the job
Give the model one job at a time. A prompt earns its length when it removes decisions the model would otherwise make for itself.
Set the instruction at the level where the model can act on it. It should be specific enough to guide behavior while leaving room for useful judgment.2 For a cold email, say whether the model is researching an account, finding a reason to contact someone, drafting a subject line, writing a body, or reviewing a draft. Each task needs its own success condition.
Use a simple opening instruction:
Draft a cold email for the recipient described below. Use only the supplied facts. Find one supported reason to contact this person, explain the likely business relevance in plain language, and end with a low-friction next step.
Then add the audience, sender context, goal, available research, and output format. Tell the model what to leave out when the input does not support a claim. That keeps a missing fact visible instead of giving the model room to fill the gap.
Load the context
Personalization needs real context to attach to. Build the input before asking for prose.
Gather the research you have on the prospect before drafting.3 Keep the notes in one working document as you collect them, so the prompt receives a coherent account picture instead of scattered fragments.4 Remove duplicate notes and separate observed facts from your own interpretation before pasting them in.
For account research, ask for a defined set of fields using public sources and require a citation for every fact. A useful request is: "Research [company]. Give me their founding, ownership, locations, service lines, certifications, and estimated headcount using only public sources. Cite each fact."5
Add a source contract beside the research request. State which sources the model may use, how recent news must be, which claims need citations, which claims are prohibited, which fields the output must contain, and how the model should express confidence. These controls belong in the prompt because they change what the model is allowed to say.6
Company context still leaves a gap when the email gives the recipient no reason to care. Ask for a reason connected to the recipient as a person, then test whether it follows from the role and the available facts.7 If the model can explain only why the company matters, send it back to the research step.
Define the sender
The sender profile gives the draft a point of view and lets you reuse the stable parts of the prompt while changing the prospect notes.
Create a sender role before asking for an email.8 Give the model enough context to write from the sender's perspective, including what the sender created, where the sender is, why it could matter to the recipient, and why the message is being sent.9
Answer these questions:
- Who is sending the message?
- What does the sender offer or know?
- Where is the sender in relation to the market or customer?
- Why could the offer matter to this recipient?
- Why is the sender contacting this person now?
The answers give the role substance without asking the model to invent a personality. Reuse the role with different prospects.10 Keep the sender profile stable and replace the account, recipient, observations, and goal for each new draft.
Set the output task
Once the input is ready, tell the model exactly what to produce. A good output instruction gives it room to write while keeping the review criteria visible.
Include instructions for the subject line, tone, and copywriting framework when those choices matter to your process.11 Put the prospect notes below a clear boundary such as "Notes to include casually in the body of the email:" so the model can distinguish instructions from source material.12
A reusable prompt can look like this:
Sender: [sender profile]
Recipient: [recipient role and company]
Goal: [the action or reply you want]
Account research: [supported observations and citations]
Task: Write a concise cold email in my tone. Use one relevant observation, connect it to a plausible business issue, and ask for the next step. Do not invent facts. If the notes do not support a personalization point, leave it out.
Output: Give me three subject line options and one email body. After the draft, list the observations used and flag any sentence that needs human verification.
When a prior conversation supplies the strongest context, add the instruction "Use my tone. Write a prospecting email for this company using the info from our last call".13 If you already wrote a draft, give it to the model as the starting material and ask for a focused revision.14 This keeps your judgment in the message while the model works on structure and wording.
Use the model for a section of the email when that is the cleanest job. It can help with an opening, a subject line, a relevance sentence, or a call to action. A prospecting workflow that uses AI for chunks of emails requires learning when to apply it and how to prompt it well.15
Test and refine
Treat the first output as a diagnostic. Read it for unsupported claims, generic relevance, awkward tone, and a call to action that does not match the goal.
Start with a minimal prompt using the best model available, then add instructions and examples in response to the failure modes you see.16 Examples give the model a reference for tone, specificity, and acceptable structure, and few-shot prompting remains a strongly recommended practice.17 Give it one approved email and one weak email if you have both, then explain what makes each useful or unusable.
When the output misses, ask the model to suggest a better prompt before adding a long list of new rules.18 Make one change at a time so you can tell which instruction fixed the problem. Keep the prompt readable enough for another rep to see what it is supposed to do.
You can also run a direct review pass:
Roast me. What could I do better in my outbound pitch?
That prompt is useful after the model has produced a draft or after you have written your own version.19 Use the response to revise the instruction, source notes, or sender profile. Do not use criticism as a substitute for checking the facts.
For repeatable outbound work, review the interaction logs, including inputs, outputs, tool calls, and actions, to find candidate rules and hard examples that can improve the prompt over time.20 Give the system recurring outcomes to work toward instead of treating every request as an isolated prompt.21 The recurring outcome might be a supported reason to contact a person, a clean first draft, or a review-ready set of alternatives.
If your research came from a conversation, feed that context into a preparation prompt before drafting. Ask for insights and questions that clarify the company's situation and the relevance of your offer.22 Then carry the strongest supported insight into the email instead of turning the whole conversation into a summary.
What not to do
These mistakes make a prompt sound complete while leaving the model without a reliable basis for the draft.
- Do not use vague, high-level guidance that assumes shared context or fails to give concrete signals for the desired output.23
- Do not let the model choose its own source rules, news limits, citation standard, prohibited claims, output fields, or confidence treatment.6
- Do not make a full email the only task when a subject line, relevance sentence, or review pass would give you a cleaner piece to inspect.15
- Do not leave the purpose of the email implicit when the notes do not state it. Tell the model what the message is meant to achieve.24