customers

This skill should be used when the user wants customer discovery work — interview guides, problem validation, personas, jobs-to-be-done, or synthesizing interview/survey feedback — e.g. "make a custom

SKILL.md

GitHub

Customer Discovery — Interviews, Personas, Synthesis

The cheapest way to avoid building something nobody wants: talk to people correctly and read the evidence honestly. Two jobs: prepare (guides, ICP hypothesis, who to talk to) and synthesize (turn raw notes into ranked, quoted insights).

Shared conventions (venture suite)

  • Workspace: locate ventures/ in the current working directory (ask once and create if missing). All outputs go to ventures/<slug>/<NN-module>/.
  • Templates are read-only. Files under a skill's references/ live inside the installed plugin and are shared by every venture. Never write to them. "Fill template X" always means: copy it into ventures/<slug>/<NN-module>/ under a dated name and fill the copy.
  • Read VENTURE.md first. When done: update its Status table and append key findings to "Key facts & numbers" (format: - [module, YYYY-MM-DD] finding (source, confidence)).
  • Language: converse in the user's language. Write formal deliverables (reports, canvases, battlecards) in the report_language set in VENTURE.md. Persian reports keep standard English business terms as-is (TAM, SAM, SOM, CAC, LTV, GTM, churn, retention...).
  • Research integrity: cite every external claim (source + date). Mark every estimate with confidence (high/med/low). Never fabricate numbers — write "no data found" instead.
  • Chain: end by updating VENTURE.md and suggesting the logical next module.

Mode A — Prepare (before/while talking to people)

1. ICP hypothesis

From VENTURE.md + riskiest assumptions, draft: who exactly has this problem worst (segment + situational trigger), where to find 10 of them this week (specific: which Telegram groups, associations, LinkedIn filters, physical locations), and what would disqualify someone as a target. Save as 01-discovery/icp-hypothesis-YYYY-MM-DD.md.

2. Interview guide — Mom Test rules enforced

Build from references/interview-guide-template.md + references/mom-test-rules.md. Non-negotiables:

  • Questions about their life, past behavior and current workarounds — never about your idea or hypotheticals ("would you use/pay...?" is banned).
  • Target the riskiest assumptions from VENTURE.md — each assumption maps to ≥1 question.
  • Interview language = customers' language. Write the guide in it (Persian guide for Iranian customers), regardless of report language.
  • Include: intro script, 8–12 core questions with follow-up probes, closing asks (referral to 2 more people; permission to follow up), and a per-interview note template (so notes arrive synthesizable).

Save as 01-discovery/interview-guide-YYYY-MM-DD.md. Coach briefly: 15–20 min per interview, record or note verbatim quotes, aim for 10–15 interviews before synthesizing, log each note file into 01-discovery/interviews/.

Mode B — Synthesize (notes/survey data exist)

1. Ingest

Read everything in 01-discovery/interviews/ (+ any survey exports, support logs, review dumps the user points to). Note n and how participants were sourced — selection bias goes in the report.

2. Extract & rank

Follow references/synthesis-template.md:

  • Pains ranked by frequency × severity — severity shown by behavior (paying for workarounds, spending hours, past attempts to solve), not politeness. Verbatim quotes attached (anonymized: "P4, clinic owner").
  • Current solutions & what they cost (money/time) — this is willingness-to-pay evidence.
  • Segments that emerged (who feels it worst — often ≠ the ICP hypothesis; say so when it diverges).
  • Personas / JTBD: only as many as the data supports (usually 1–2 from 10–15 interviews). Format: "When <situation>, I want <motivation>, so I can <outcome>" + hire/fire criteria for solutions.
  • Disconfirming evidence — mandatory section: who didn't care, which assumptions cracked. If everyone loved everything, suspect the questions (or the sample) and say so.

3. Verdict on assumptions

For each riskiest assumption in VENTURE.md: validated / refuted / still unknown + the evidence line. Update VENTURE.md accordingly (check off, revise, or add new ones discovered).

Save as 01-discovery/synthesis-YYYY-MM-DD.md in report_language (quotes stay bilingual: original + translation).

Close the loop

Update VENTURE.md (status, key facts: top pain + quote, emergent segment, willingness-to-pay evidence). Suggest next: assumptions validated → /venture:market-research (size it) or /venture:model; refuted → reframe with /venture:start; product live → feed into /venture:pmf.

Quality bar

  • Compliments are not data. "I'd definitely use this" without past-behavior evidence = polite rejection — count it as nothing.
  • n and sampling method appear in every synthesis; 5 friendly interviews ≠ validation, and the report says so.
  • Insights without verbatim quotes don't ship.