founder-os-mom-test-auditor

FounderOS Mom Test interview auditor. Splits a customer interview transcript into founder and interviewee turns, has Laya tag each turn (compliment, hypothetical, pitching, past behaviour, commitment

SKILL.md

GitHub

Mom Test interview auditor

This is step 7 of interviews-without-fooling-yourself: the line-by-line audit. Laya pre-tags every turn. You confirm the tags, and you write the feedback. The Mom Test is Rob Fitzpatrick's; use his terms and point the founder at his book for the full argument.

When to run

  • The founder has a transcript or detailed notes from one interview, ideally within a day of the call.
  • Save it first, so labels and audits have a row to hang on: add_interview(founder_id, interviewee=<pseudonym>, notes=<transcript>, conducted_on=<date>). interviewee is a label like studio-owner-3, never a name or contact. Set commitment only if the transcript plainly shows it; otherwise leave it none and record the audit's view in the audit findings.
  • The transcript should be Speaker: text lines. If it is prose notes, ask the founder to mark who said what, or run it anyway and tag it yourself.

Laya pass

python3 "$FOUNDER_OS_DIR/scripts/audit.py" mom-test \
  --founder-id <founder_id> --interview-id <interview_id> \
  --founder "<founder's speaker label>" --lang <en|pt|...> --record

Always pass --lang; it picks the checkpoint. The output is JSON:

  • turns[]: n, speaker, role (founder or interviewee), text, tags (per tag value, p, trust), patterns and label_ids.
    • patterns.bare_yes_no: the answer opened with a bare yes or no. Mark it doubtful; the question before it was probably leading.
    • patterns.closed_question: the founder asked a yes/no question.
    • patterns.hypothetical_wording: the interviewee said would, maybe, probably, everyone, people and the like.
  • summary: turn numbers per tag, founder_talk_share, pitched_before_first_fact, facts_vs_struck, and the commitments:
    • commitment_candidates: commitment tags on turns that are not hypothetical, by time, reputation (an intro) and money.
    • commitment_talk_only: commitment tags on hypothetical turns. "I'd pay twenty euros for that" is talk about paying, not a payment. Never count these.
    • suggested_commitment: the strongest candidate, or none.

Every tag in this set has trust: unmeasured. No accuracy has been measured on it yet. Read each flagged turn yourself before you repeat Laya's call, and look for turns it missed. In a first live run on the english checkpoint, commitment_money fired on almost any turn that mentioned money, and past_behaviour missed most dated incidents. Check both every time.

Claude pass

Write the report in this order, citing turn numbers for every claim ("turn 4", "turns 6-8"):

  1. Verdict in one line. Could anything in this interview have changed what they build? If not, say so.
  2. Facts that survived. Dated incidents, money or hours already spent, tools in use. Quote the short phrase and give the turn.
  3. Struck and flagged. Compliments (struck, carry no information), hypotheticals and generic claims (flagged fluff), bare yes/no answers (doubtful). Count facts against struck lines.
  4. The founder's side. Pitching, especially before the first fact; leading or future-tense questions, each with a rewrite about the recent past; talk share above about a third.
  5. Commitment. What the interviewee agreed to do next and what it costs them: time, reputation (an intro) or money. "Keep me posted" is none.
  6. Next interview. Two or three concrete changes.

Do not soften it. A warm conversation with no facts is a failed interview, and the founder is better off hearing it now.

Without Laya

If mode is claude-only, degraded_reason says why. The turns are split and numbered and the patterns are filled; tags is null. Tag each turn yourself with the same tags, using the question text in laya/mom-test.json, and write the same report. Say in one line that Laya was unavailable. Write no labels: labels are Laya's judgments, and there are none.

Label capture

After the report, show the founder the tags you disagreed with and ask them to confirm. Then record every tag they accepted or rejected:

python3 "$FOUNDER_OS_DIR/scripts/labels.py" accept --founder-id <id> --by <founder_id> <label_id> ...
python3 "$FOUNDER_OS_DIR/scripts/labels.py" reject --founder-id <id> --by <founder_id> <label_id>

A rejected yes/no tag flips. When the founder says "that's right" about the report as a whole, accept the tags the report relied on. Tags nobody reviewed stay unreviewed; never accept them in bulk to fill the set. Each reviewed tag is a fine-tuning example. The cross-founder export is operator-only and is never run from a founder session.

Ledger writes

  • The interviews row, before the audit (see When to run).
  • Labels, one per turn and tag, from audit.py --record; corrections from labels.py.
  • One audits row with your verdict: record_audit.py --table interviews --id <interview_id> --auditor claude --check mom_test --verdict pass|flag|fail --findings "<one paragraph with turn numbers>".
  • Never change commitment or earlyvangelist on the interview from here. The interview is written once; suggested_commitment goes in the audit findings, and the earlyvangelist call belongs to founder-os-earlyvangelist-qualifier.