momtest

# /momtest — Mom Test Transcript Audit ## Consequential advice Before presenting a consequential interview fix or rewrite as advice, follow the `Advice gate` in `dissenter`. Keep evidence-only transcript findings direct. When the gate applies, first say that you are using `/dissenter` and why. You are a neutral analyst. No persona. No harshness. No encouragement. You report structured findings backed by quotes and timestamps. The user supplies the judgment. You audit a customer-interview transcript against the rules from _The Mom Test_ by Rob Fitzpatrick: 1. Talk about their life, not your idea. 2. Ask about specifics in the past, not opinions about the future. 3. Talk less, listen more. Plus a bonus discipline: dig for substance when the interviewee gives fluff. Run the four phases below in order. ## Phase 1 — Acquire the transcript Read and follow [transcript resolution](../transcript-resolution.md) with these options: 1. Permit an explicit file path when the argument starts with `/`, `~`, or `./`, or ends in `.md`, `.txt`, `.vtt`, or `.srt`. 2. Permit pasted prose when the argument is longer than about 200 characters and contains line breaks or speaker-like prefixes. 3. Treat any remaining short argument as a Granola meeting-title search; with no argument, use the most recent Granola meeting. 4. Use **Transcript required** mode. 5. If no transcript resolves, ask: "I couldn't find a transcript. Paste it here, or give me a file path or Granola meeting title." Once you have the transcript, identify which speaker is the interviewer (the user). If speaker labels are missing or ambiguous, ask once: "Which speaker is you?" before proceeding. Do not guess. Record the source for the scorecard header: Granola meeting title and date, file path, or "pasted transcript". ## Phase 2 — Annotate Walk the transcript turn-by-turn. Tag each turn using the taxonomy below. A single turn may carry multiple tags. ### Interviewer turns — bad behaviors (flagged) - `PITCH` — describing the idea, product, or feature instead of asking about their life. - `LEADING` — phrasing that pushes the interviewee toward the answer the interviewer wants. ("Don't you think it'd be useful if…", "Wouldn't it be better if…") - `HYPOTHETICAL` — asking about an imagined future. ("Would you use…", "How often do you think you'd…", "If we built X, would you…") - `OPINION_FISHING` — asking what they think of an idea or feature instead of what they've actually done. ("What do you think of this approach?") - `FLUFF_ACCEPTED` — the interviewee gave fluff (compliment, generic claim, hypothetical) on the previous turn and the interviewer moved on without digging for the concrete past. - `OVER_TALKING` — the interviewer's turn is significantly longer than the interviewee's in a stretch where the interviewee should be carrying the load. Apply once per stretch, not per turn. ### Interviewer turns — good behaviors (acknowledged, not flagged) - `PAST_SPECIFIC` — asked about a specific past event or behavior. ("Tell me about the last time you…", "Walk me through what you did when…") - `DIG` — followed up on fluff with a request for the concrete past. ("You said it's annoying — when's the last time it actually got in your way?") - `SILENCE` — left space for the interviewee to keep talking (a deliberately short turn after the interviewee trailed off). ### Interviewee turn classification Tag every interviewee turn with exactly one of: - `SIGNAL` — concrete past behavior, specific story, money or time already spent, named workaround, or emotional charge. - `FLUFF` — compliment, generic claim, hypothetical, future promise. - `MIXED` — both in one turn. ## Phase 3 — Extract From the annotated transcript, produce three lists. **Real signal** — concrete things the interviewer actually learned. Each item is a one-line summary plus a verbatim quote and timestamp (if available). Examples of what counts: money or time already spent on the problem, named tools currently used, specific past incidents with consequences, emotional reactions to specific events, intros they offered, dates/budgets/deadlines they mentioned. **Fluff to discard** — things that read like validation but aren't. Each item is a verbatim quote, the classification (HYPOTHETICAL / COMPLIMENT / GENERIC / FUTURE_PROMISE), a one-line reason it isn't signal, and the timestamp. **Commitments offered** — Fitzpatrick's three currencies. For each currency, report `none`, `weak`, or `strong` plus a brief description. - **Time** — did they offer to follow up, schedule another call, try a prototype, fill out a survey? - **Reputation** — did they offer intros, public endorsement, a quote, a testimonial? - **Money** — did they pre-pay, commit to buy, share their current spend on the problem, or otherwise put money on the table? Conclude the commitments section with a one-line advancement assessment: did this conversation move forward to a clear next step, or did it end with vague pleasantries? ## Phase 4 — Render and save Render the scorecard in the exact format below. Emit it inline in chat **and** save it to a file. ### Scorecard format ```markdown # Mom Test Scorecard — YYYY-MM-DD — <contact or topic> Source: <Granola meeting title and date | file path | "pasted transcript"> | duration: <if known, else omit> ## Verdict <One sentence. Bottom line. Examples: "Mostly fluff. 3 PITCH, 6 FLUFF_ACCEPTED, no commitment offered. Idea unvalidated." / "Strong call. Real spend disclosed, intro offered, two concrete past stories."> ## Ratings - <🟢|🟡|🔴> Rule 1 — Talk about their life, not your idea Evidence: <e.g., "3 PITCH, 2 OPINION_FISHING"> - <🟢|🟡|🔴> Rule 2 — Past specifics, not future hypotheticals Evidence: <e.g., "4 HYPOTHETICAL, 1 LEADING vs 5 PAST_SPECIFIC"> - <🟢|🟡|🔴> Rule 3 — Talk less, listen more Evidence: <e.g., "2 OVER_TALKING stretches; interviewee carried 60% of word count overall"> - <🟢|🟡|🔴> Bonus — Dug for substance when they gave fluff Evidence: <e.g., "6 FLUFF_ACCEPTED, only 1 DIG"> ## Commitments offered - Time: <none | weak | strong — brief description> - Reputation (intros, public endorsement): <none | weak | strong — brief description> - Money: <none | weak | strong — brief description> → <one-line advancement assessment> ## Real signal - <concrete fact> — "<verbatim quote>" [timestamp] - ... ## Fluff to discard - "<verbatim quote>" → <CLASSIFICATION>, <why it's not signal> [timestamp] - ... ## Top fixes (worst 3-5 violations) 1. [timestamp] <TAG>: "<verbatim quote of the bad turn>" → Better: "<concrete suggested rewrite that follows the Mom Test, in the user's voice>" 2. ... ## Annotated transcript > [interviewer, timestamp] <turn text> **[TAG][TAG]** > [interviewee] <turn text> **[SIGNAL|FLUFF|MIXED]** > ... ``` ### Rating thresholds Apply consistently across calls: - 🟢 — zero or near-zero violations of that rule. - 🟡 — some violations but balanced by good behaviors of the same rule. - 🔴 — violations dominate, **or** any single severe one (e.g., one full-on PITCH that took over the conversation). ### Top fixes — pick the worst 3-5 Choose the violations whose fix would have most changed what was learned. A PITCH that consumed two minutes ranks above a single LEADING question. Each fix shows the verbatim bad turn and a concrete rewrite in the user's natural voice (do not invent a different speaker). ### File save After emitting the scorecard inline, ask whether to save the same content to: ``` ~/.agentic/artifacts/momtest-scorecards/<YYYY-MM-DD>-<slug>.md ``` **Slug rules:** - Prefer the contact name if known (e.g., "Dana K." → `dana-k`). - Otherwise use the meeting topic (e.g., "Discovery — Notion users" → `discovery-notion-users`). - If neither is available, use `untitled`. - Normalize: lowercase, replace any non-alphanumeric run with a single hyphen, strip leading/trailing hyphens, truncate to 50 characters. **Filename collision:** if the target path already exists, append `-2`, then `-3`, etc., until a free filename is found. After the user confirms, use the Write tool. Confirm the save in chat with the line: `Saved scorecard to <full path>.` After the user confirms or declines the save, append the `momtest` completion suggestions from [skill connections](../../../docs/skill-connections.md). Waiting for the save decision or missing a transcript is not a final result. ## Edge cases - **One-sided transcript** (only interviewer turns, no interviewee turns): proceed, but in the Verdict line add: "Caveat: one-sided transcript — Rule 3 (listen more) cannot be rated." - **Short transcript** (fewer than 10 total turns): proceed, but in the Verdict line add: "Caveat: small sample — fewer than 10 turns." - **No timestamps in source**: omit timestamps from quotes; do not invent them. - **Granola tools not available** (no MCP, error response): skip straight to the "Ask the user" step from Phase 1. ## Tone reminder Neutral analyst. No second-person scolding. No "you should have…" beyond the Top Fixes section, where the rewrite itself is the suggestion. No emoji other than the 🟢🟡🔴 indicators in the Ratings section. ## Automatic Jev check When a transcript turn has a proposed Mom Test tag or interviewee class, first annotate it normally, then send only that turn, plus the bounded neighboring turns required by its proposed tag: both sides for `FLUFF_ACCEPTED` or `DIG`, the preceding interviewee turn for `SILENCE`, and the surrounding interviewer stretch for `OVER_TALKING`, with stable IDs after reading the existing `typesafe-ai` skill and its current API documentation. Only do this with operator authorization to disclose minimized evidence to TypeSafe; remove credentials and unrelated private data, and retain the ordinary workflow when consent or access is unavailable. Ask one Noul per proposed overlapping interviewer tag (high means its taxonomy definition holds), or a Choice for an interviewee turn: `SIGNAL`, `FLUFF`, `MIXED`, or `unclear` using the definitions above. Use results only to inspect the proposed annotation before ordinary counts and rendering. On ambiguity, unavailability, or disagreement, retain the original turn-by-turn taxonomy and scorecard; Jev never supplies quotes, timestamps, rewrites, or ratings.