AI prompt template
AI prompt to build an ideal customer profile
An ICP written from opinions is a guess. This prompt starts from the public pages of customers you won and lost, and states only what the evidence supports.
When to use it
- You have at least five customers you won and a few you lost.
- You are writing the criteria for a list build.
- You want something to challenge your own assumptions.
The prompt
Paste it into the model you use, then replace each bracketed input.
We sell: [WHAT_WE_SELL] Below are public pages of companies that bought from us (WON) and ones that did not (LOST). Draft an ideal customer profile using only these pages; never invent a pattern. WON: [WON] LOST: [LOST] Return: 1. Traits shared by most WON companies that LOST companies lack. Each with the companies it appears in and one verbatim quote. 2. Traits shared by both groups (so they do not predict anything). 3. Roles named on the WON companies’ pages as owners of the problem we solve, if any; otherwise "not found". 4. Three criteria a web page could show, written so they can be checked. 5. What sample size limits your confidence. If the pages are too few to support a trait, say so and write "not found".
| Input | What to put there |
|---|---|
[WON] | Pages of customers you won, each headed by the company name and address. |
[LOST] | Pages of companies you lost or that never converted. |
[WHAT_WE_SELL] | One sentence on your product. |
Why it works
- Comparing won and lost is what makes it an ICP rather than a description of customers.
- Item 2 catches traits that look important but appear everywhere.
- Item 5 forces the model to say when the sample is thin.
An example
Illustrative. The companies and pages below are invented.
Input
WHAT_WE_SELL: dispatch software for plumbing firms WON: four plumbing firms LOST: three electrical firms (illustrative)
Output
1. Four of four WON list emergency call-outs on their homepage, for example "24 hour emergency plumbing" (WON1). No LOST firm does. 2. Both groups list service areas by town. 3. Roles: not found. 4. Criteria: lists emergency call-outs; lists service areas by town; has a booking page. 5. Seven companies is a small sample; treat the first trait as a hypothesis. (Illustrative output on invented companies.)
Variants
Only from won customers
We sell: [WHAT_WE_SELL]. Using only the pages below, list what the companies have in common, each trait with at least two companies and a verbatim quote. Never invent a trait. If fewer than two companies show a trait, do not list it. [WON]
Segments
Using only the pages below, group the companies into at most three segments by something a web page shows. For each: members, the defining trait with a quote, and "not found" for anything you cannot support. Never invent a segment. [WON]
Limits
What are the limits of the draft an icp from your best customers prompt?
Related
Qualify a company against your ICP
A prompt that scores a company against your written ICP from its pages only, with a quote per criterion and an explicit “unknown” where it cannot tell.
Pick the right buyer from public titles
A prompt that picks who to write to from the people and roles a company’s own pages name, and says “not found” when no page supports a choice.
Research an account from its own pages
A prompt that turns a prospect’s own pages into sourced notes for a cold email: what they do, what changed, who owns the problem, each with a quote.
Cold email templates
Templates by situation and by role.
Research and writing, with a quote behind every fact
Outlevo runs these steps for each company it finds, and shows the page each fact came from.