ideal-customer-profile

---
name: ideal-customer-profile
description: Turn your last 20-50 won and lost deals into a Green, Yellow and Red ideal customer profile, back-tested on real customers; use when asked who to target or why nobody replies.
---

# Find out who actually buys from you

Filter rules from your own wins and losses, proven on real customers.

## Claude does

1. Collect the last 20-50 closed-won and closed-lost deals (CRM or memory): the seven row fields, who signed, who first engaged, months to close, ARR, months as customer, expanded or churned, loss reason. Under 20: `hypothesis`, rebuilt at 20; else `evidence`.
2. Exclude customers under 90 days, one-off mega deals and churned accounts; rank the rest by post-sale value (top 20% ARR, NRR above 115%, lowest churn, fastest time to value): the best customers.
3. Row by row (industry, size, geography, decision maker, dependent technology, use case, current platform), nail Green first: what best customers share that losses do not; signal is 60% of them across 5+ deals, else `unknown`.
4. Then Yellow (inbound only) and Red (disqualify even inbound) from the losses, by loss reason.
5. Cells are ten-second filters: "50-500 employees", not "mid-market".
6. Ask "how do you know that? where would you check that?" per cell; tag deal IDs; mark Priority on the 2-3 go/no-go rows.
7. Name the best customers' trigger, 2-3 channels that reach them (or which is missing), 2-5 back-test customers including one surprise, and the ten-second sentence about the perfect prospect.

## Then Codex does

Codex builds the deal table, worksheet and back-test sorting every deal and named customer through the rules, all mechanical. Fill {COMPANY}, {STATUS}, {DEALS}, {BEST_IDS}, {GRID}, {TRIGGER_AND_CHANNELS}, {TEST_CUSTOMERS}, {TEN_SECOND_SENTENCE} into CODEX.md, then run in the working folder:

```
codex exec --sandbox danger-full-access --skip-git-repo-check -C "<working folder>" - < CODEX.md
```

## Claude checks

1. `DEALS.csv` holds exactly the deals supplied.
2. `BACKTEST.md`: test customers and last 10 wins land Green, last 10 losses Yellow or Red; a great customer sorted otherwise: change the cell, not the customer.
3. Green creep: a row matching over half of won deals discriminates nothing.
4. Red cells name a cost and deal IDs; every cell is a filter; unsourced cells stay `unknown`; losses split by reason.

Any failure: fix, rerun, recheck; show nothing until all pass.

## Rules

- Every cell cites deal IDs or says `unknown`; never invent a filler.
- One page; end with the ten-second sentence.
- Red must cost something: churn, refunds, support burden; else Yellow.
- No vague bands; an ICP that disqualifies nobody is not an ICP.
- Never average across loss reasons or segments, or refine from new, churned or one-off mega-deal customers.
- No trigger, no reachable channel, "companies that need X", no rubric: fail.
- Reviews skew to power users, tickets to problems; weigh that.
- Never ship without the back-test; refresh quarterly or at 10 new customers, 90 days is overdue; never loosen under sales pressure.

## If Codex is not installed

Claude does that half and says so.

## Built from

Stars: api.github.com.

- Stage 2 Capital ICP-Builder, https://github.com/stage-2-capital/ICP-Builder (1 star; Mark Roberge's GTM fund)
- borghei/Claude-Skills, https://github.com/borghei/Claude-Skills/blob/main/project-management/gtm/ideal-customer-profile/SKILL.md (702 stars)
- alirezarezvani/claude-skills, https://github.com/alirezarezvani/claude-skills/blob/main/marketing-skill/skills/marketing-strategy-pmm/SKILL.md (25,542 stars)
- tech-leads-club/agent-skills, https://github.com/tech-leads-club/agent-skills/blob/main/packages/skills-catalog/skills/(gtm)/positioning-icp/SKILL.md (5,121 stars)
- coreyhaines31/marketingskills, https://github.com/coreyhaines31/marketingskills/blob/main/skills/customer-research/SKILL.md (46,908 stars)
- phuryn/pm-skills, https://github.com/phuryn/pm-skills/blob/main/pm-go-to-market/skills/ideal-customer-profile/SKILL.md (26,000 stars)

# Find out who actually buys from you - Codex task

Claude has decided everything; turn it into files in the current folder using nothing beyond the inputs.

## You are given

- {COMPANY}: the company.
- {STATUS}: `evidence` or `hypothesis`.
- {DEALS}: 20-50 closed-won and closed-lost deals with the fields in the CSV columns below; fields may be missing.
- {BEST_IDS}: the won deals ranked as best customers.
- {GRID}: seven rows, one per CSV column from industry to current_platform, each with Green, Yellow and Red filter rules, Priority yes or no, source deal IDs or `unknown`.
- {TRIGGER_AND_CHANNELS}: the buying trigger and 2-3 channels.
- {TEST_CUSTOMERS}: 2-5 real customers with fields and the bucket each actually ended up in.
- {TEN_SECOND_SENTENCE}: verbatim.

## Produce

1. `DEALS.csv`: one row per deal, columns `id, outcome, company, industry, headcount, geography, decision_maker, dependent_tech, use_case, current_platform, months_to_close, arr, months_customer, expansion_or_churn, loss_reason, best`. IDs W1 onwards for won, L1 onwards for lost, in the order given; `best` is yes for {BEST_IDS}; copy fields as given, blank if missing.

2. `ICP-WORKSHEET.md`, under 400 words: `# Ideal customer profile - {COMPANY}`; if {STATUS} is `hypothesis`, the line `**Hypothesis, not evidence. Rebuild at 20 deals.**`; a table with columns Row, Green (target outbound), Yellow (inbound only), Red (disqualify), Priority, Source, one line per {GRID} row, cells verbatim; `## Trigger and channels` ({TRIGGER_AND_CHANNELS}); `## Refresh by` and the date three months from today; last line {TEN_SECOND_SENTENCE}, unchanged, nothing after it.

3. `BACKTEST.md`, four sections:
   - **Tally**: per row, how many best, other won and lost deals match its Green rule, as counts and percentage of each group.
   - **Sort**: run every deal and test customer through {GRID}: Red if any row matches its Red rule, else Yellow if any matches Yellow, else Green, skipping rows whose cell is `unknown` or field blank. Table: id, name, predicted, actual, hit or miss.
   - **Score**: hit rate for the last 10 wins, last 10 losses and test customers, separately; list every miss with the row that caused it.
   - **Losses by reason**: lost deals per stated reason (price, fit, timing, other), never merged.

## Rules

- Use only the inputs; invent no deal, rule, figure or name; `unknown` stays `unknown`.
- Apply rules literally; never loosen one to make a customer fit, report the miss.
- Every number must be recomputable from `DEALS.csv`.
- British English, no em dashes.

## Return

Print only: each file's path; `ICP-WORKSHEET.md` word count; deals as won / lost / best; per row, percentage of won deals matching Green; hit rates for last 10 wins, last 10 losses, test customers; every miss with its row; cells still `unknown`. Then stop.