An AI cold email prompt needs a job for the model, material to work from, and boundaries for what may appear in the draft. Keep these pieces separate so a weak email can be traced to thin research, missing sender context, or loose instructions. Give AI real observations, experiences, and context, then ask it to fit them into a structure shaped by account research. The human supplies the substance; AI handles expression.1 Start with the smallest prompt that can do the task. Add an instruction or example only when a failure shows what is missing.2
Start with the account brief
Start research before copy. Give AI enough company and person context to find a reason for the email, and keep useful material available for drafting.
Begin by asking AI to research the account and find relevant company information.3 A useful research instruction is:
"Research [company]. Give me their founding, ownership, locations, service lines, certifications, and estimated headcount using only public sources. Cite each fact."4
Use a second pass to turn company facts into possible challenges. This prompt gives the model a business situation, the framework you want to discuss, and the company resources it should use:
"I'm going to engage with [customer name] from [company name] on how [my framework] could apply to their business. Based on the resources you can pull on their company, what challenges might they be facing? [included company website]"5
Use a company-level trigger to open the research, then find a relevant reason to contact the person that fits the human and guides the email.6 Move on when the brief has usable facts, a plausible problem to explore, and a reason the recipient should care.
Define the sender
AI needs a clear point of view to produce an email that sounds like you. Keep that reusable block separate from each prospect's notes.
Describe who is sending the email, what they created, where they are, why it could matter to the recipient, and why they are writing.7 This gives AI the material to write from your perspective.8 Reuse the completed role block with different prospects.9
Use plain details about your offer, your experience, and why you have earned the right to raise the topic. Leave prospect-specific facts out of this block so the role stays stable while the account brief changes.
Specify the output
Tell AI what it must produce. A narrow output makes it easier to tell whether a problem came from the prompt or the account notes.
Set the email task together with instructions for the subject line, tone, and copywriting framework.10 A useful starting instruction is:
"Use my tone. Write a prospecting email for this company using the info from our last call."11
Paste the sender role at the top of the prompt.12 Put the prospect notes below a clear boundary such as:
"Notes to include casually in the body of the email:"13
Ask for the exact format you will review, such as the subject line, email body, and any claims that need checking. Keep the task focused on the draft you need now. Combining research, analysis, positioning, and copy in one undivided instruction makes it harder to see which part failed.
Add the prompt contract
A prompt contract tells AI what it may use, what it must show, and where it must stop. Write these rules before polishing the wording of the email.
Set the approved sources, maximum news age, required citations, prohibited claims, output fields, and confidence rules in the prompt.14 Ask AI to mark missing information and low-confidence conclusions instead of filling gaps with plausible language.
Leave room for useful judgment. An effective prompt is specific enough to guide behavior and flexible enough to give the model strong heuristics.15 Use constraints that make the result easy to inspect, such as requiring each account fact to retain its source and separating observed facts from suggested angles.
Test and improve the prompt
Treat the prompt as a working draft. Test it against a real account, inspect the failure, and change the smallest part that explains the problem.
Add examples when the output keeps missing the shape or tone you want. Few-shot prompting is a strongly recommended practice.16 Choose examples that show the desired behavior, such as turning a sourced company fact into a restrained observation in an opening line.
You can also ask AI to suggest a better prompt when the current instruction produces weak results.17 Ask it to identify the instruction that caused the failure, then decide whether the change belongs in the role, account brief, output task, or contract.
Once the prompt runs repeatedly, retain the inputs, outputs, tool calls, and actions. Those records can reveal hard examples and candidate rules that improve the prompt over time.18 Move on after the same failure no longer appears across the accounts you use. One good draft is not enough.
Draft from human material
Give AI material with a point of view before asking it to write. This keeps the email grounded in what you learned and gives you something concrete to inspect.
Start with a human-written draft as the working material for AI.19 Add the observations from the account, what you heard in the last conversation, and the reason you want the recipient to reply. Ask AI to tighten the language, preserve the substance, and fit the material to the requested tone.
For an account with a possible in-house objection, use a separate prompt to anticipate and prepare for that DIY objection.20 Keep objection preparation separate from the opening email so the draft does not become a catalogue of every possible response.
Review the draft
Use AI as a critic after it produces the draft. Give it a clear standard and ask for changes you can act on.
A direct review prompt is:
"Roast me. What could I do better in my outbound pitch?"21
Follow it with the email, the account notes, and the prompt contract. Ask the reviewer to point to the line that needs work, explain what the available account material supports, and propose a replacement that stays within the stated rules. Keep the final decision with the person who knows whether the observation is true.
What not to do
These failures make a prompt sound complete while leaving AI without a reliable way to choose what belongs in the email.