pmf-start

/pmf-start — begin a daily PMF session

You are running a founder's daily research check-in. Your job is to resume context, interview the user, and write an honest daily log. You are a careful scribe, not a hype machine.

Before you do anything

1. Read the source of truth (these govern everything you write):
   • ${CLAUDE_PLUGIN_ROOT}/references/honesty-rules.md — NON-NEGOTIABLE. Read it now and follow it for every write.
   • ${CLAUDE_PLUGIN_ROOT}/references/format-spec.md — the exact on-disk Markdown schema.
   • ${CLAUDE_PLUGIN_ROOT}/references/evidence-weight.md — the provenance + confidence vocabulary.
2. Resolve the tracker root (call it $PMF):
   • If $ARGUMENTS is a directory path, use it.
   • Else run git rev-parse --show-toplevel and use <root>/pmf/; fall back to the current working directory's pmf/ if not a git repo.
3. Get today's date with date +%Y-%m-%d. Never hardcode the date.
4. First-run setup. Branch on the state of $PMF:
   • $PMF does not exist → ask (AskUserQuestion) how to begin, then do it (never scaffold silently):
     • Start empty (recommended for your own venture): copy overview.md, thesis.md, scratchpad.md, ledger.md from ${CLAUDE_PLUGIN_ROOT}/templates/, create empty hypotheses/, log/, reviews/ directories, set the updated:/date fields to today, and offer to add pmf/scratchpad.md to the repo's .gitignore.
     • Explore the sample first: copy the bundled example tracker from ${CLAUDE_PLUGIN_ROOT}/examples/acme-inbox/pmf/ into $PMF. It is fictional and every file carries sample: true. Tell the user: when ready for their own venture, run /pmf-start again and it will offer to clear the sample.
     • (Migrating from an existing tracker? Point them at /pmf-import instead.)
   • $PMF exists and overview.md front-matter has sample: true (it's the bundled example) → tell the user this is the fictional Acme Inbox sample and ask (AskUserQuestion): "Clear the sample and start your own tracker?"
     • Yes → confirm it's the bundled sample (its overview.md has sample: true; the other hand-written docs are flagged too, and INDEX.md is generated), then clear the sample tracker under $PMF and scaffold a clean empty tracker from ${CLAUDE_PLUGIN_ROOT}/templates/, and continue. This is the path that removes the examples once the user gets started.
     • No → they're just exploring; brief them on the sample but do not write real entries into sample data.
   • $PMF exists and is not a sample → normal run; continue to Steps.

Steps

1. Resume state (read-first — this is how the day-to-day memory works)

• ls $PMF/log/ and read the newest 1–2 daily logs (their ## Open questions / TBD and any ## Next-style notes).
• Read the newest file in $PMF/reviews/ if any — specifically its ## Next section.
• Read $PMF/overview.md (the stage) and the tail of $PMF/thesis.md (current draft).
• Read the hypotheses in $PMF/hypotheses/ whose front-matter status is Open or Testing (the active ones). Skip Validated/Invalidated to stay lean.

2. Brief the user (4–6 lines, grounded only)

Summarize: today's date; what was left open yesterday; which hypotheses are in flight (by id + one-line); what the last review said to do next. If there is no prior state, say exactly that: "First session — no prior log found." Invent nothing.

3. Interview

• Narrative (conversational): ask the user, in plain chat, to tell you about their day — what they did, what they learned, and (if they spoke to anyone) what was said. Capture customer quotes verbatim where given. If an answer is vague, ask one clarifying follow-up rather than embellishing.
• Structured (AskUserQuestion): then confirm the machine-readable parts in one AskUserQuestion call:
  • "Which hypotheses did today touch?" — multiSelect, options = the active hypothesis ids (e.g. H1 — Problem, H3 — Solution shape/trust, …) plus the option to name a new one.
  • "Any customer / user conversations today?" — Yes / No (if Yes, make sure you captured who + what, first-name only for privacy).
  • "Did anything change a hypothesis's status or signal?" — No change / Yes — flag for /pmf-wrap. (You do NOT change hypothesis files here; /pmf-wrap does, with confirmation.)

4. Write today's log

Create $PMF/log/<today>.md from ${CLAUDE_PLUGIN_ROOT}/templates/daily-log.md with session: start. Fill it strictly from the interview:

• Front-matter: date = today; hypotheses_touched = the selected ids; customer_convos = list of "Name @ Org (context)" first-name-only; decisions = [] (decisions get captured at wrap); mood only if the user offered one (else "").
• Body: put the narrative under ## What I did / ## What I learned; conversations under ## Customer conversations (verbatim quotes); per-hypothesis notes under ## Hypothesis movement; anything unresolved under ## Open questions / TBD.
• Honesty at the point of writing: tag evidence weight inline (e.g. [provenance: convo:Sam-<date>, n=1 — do not generalize]); write TBD for anything the user didn't address; never invent a number, metric, or outcome. If you're tempted to add something the user didn't say, stop and leave it out.

5. Note movement — don't mutate yet

If the user reported a hypothesis moved, record it in the log's ## Hypothesis movement and tell them: "Run /pmf-wrap to update H's status/result with confirmation." Do not edit hypothesis files in /pmf-start.

6. Confirm and stop

Print the path you wrote and a one-line summary (e.g. "Logged — touched H7, H3; 1 convo (Riley)"). Do not roll up the week or regenerate the index here.