cold-start — Founding brief for a zero-data project

Start a research project from literally nothing: no transcripts, no reports, an empty Inbox/. Interview the founder about their startup, capture what's needed for reliable desk research, and convert their beliefs into testable L1 assumptions — without ever letting a founder's declaration masquerade as a finding.

Language: facilitate in the user's language; the files you write follow the repo's conventions.

Trigger

/cold-start        ← the only way in. Never auto-run, never suggest-and-start.

• This is a consciously invoked option. The SessionStart hook may mention it exists when the project has zero data — mentioning is the ceiling; the user must type it.
• Requires nothing: works with a completely empty project. If real Signals/Transcripts already exist, say so and ask whether they really want the founding-brief flow (it can still be useful to backfill Product Context.md and surface untested assumptions).

The one rule that governs everything

The founder is a primary source about the product, and an assumption source about the users.

Never create a Signals/ or Evidence/ file from a founder's declaration about users. A Signal exists only after a real interview/test (/extract-findings); Evidence only with a real, linkable external source. This is what keeps the graph honest when it's one week old.

Facilitation rules

Same style as /persona-workshop: one question at a time, reflect each answer in one sentence before the next question, probe vague answers, don't rush. Two additions:

1. Label assumptions out loud. When the answer is a belief about users ("our target group is freelance designers"), reflect it back explicitly as an assumption: "Recording that as an assumption to validate — not a finding." The founder should hear the epistemic status of their own words.
2. Hunt for the falsifiable core. Every belief gets the silent follow-up: how would we know if this is wrong? That phrasing becomes the hypothesis/backlog entry.

Question protocol

Ask in order; skip what the founder already volunteered. Reflect, then move on.

1. What is it? — the project in 1–2 sentences; stage (idea / prototype / MVP / live); what exists today that a user could touch.
2. Value proposition — what problem, for whom, why now, what's different from the obvious alternative. Ask for the exact sentence they'd say to a customer — verbatim, it's their positioning claim (an assumption, not data).
3. Existing materials — pitch deck, landing page, one-pager, competitor teardown? Ask them to drop files into Inbox/. Read them as founder claims: any external stat cited inside (market size, churn benchmarks) is re-verified at its original source before it may become Evidence/; unverifiable numbers stay founder assumptions.
4. Target group — who do they think it's for; segments and priority; who it is explicitly NOT for. ⚠️ This is the answer most likely to be treated as fact by everyone in the room — it never is. It becomes hypotheses and recruiting criteria, nothing more.
5. How do they know the problem is real? — own experience, past conversations, support tickets at a previous job? Anecdotes are leads for the backlog, not Signals — a remembered conversation can't be quoted verbatim, so it can't be data. If they ran actual interviews before, those transcripts go to Inbox/ → /extract-findings (the normal pipeline takes over for that part).
6. Alternatives & competitors — named products and non-consumption ("Excel and a notebook"). Each named competitor → a Competitors/ stub with needs_research: true, then follow the standard convention: ask whether to run initial desk research now, offering a 1/2/3-year lookback window. While here, propose 3–6 compare_categories for Product Context.md — the axes the founder believes decide wins (their pick of axes is itself informative); no vanity metrics.
7. Business model & pricing assumption — how it makes money; what they believe users will pay. Pure L1.
8. The riskiest assumption — "If one of your beliefs turned out false and killed the startup, which one?" That's the first hypothesis to test and the top of the research backlog.
9. Where do these users live? — channels, communities, forums, conferences. Feeds both desk-research sources and interview recruiting.

Outputs — each saved only with the user's approval

1. Product Context.md — fill What the product is from the founder's product facts (stage, what exists, pricing model). Add a section:

   ## Founding assumptions (L1 — founder-declared, to validate)
   <!-- Everything below is what the founder BELIEVES, recorded on <date>.
        Not findings. Each links to its Hypothesis / backlog entry.
        /persona-talk and /feature-panel must not cite these as facts. -->
   - Target group: … → [Hypothesis: …](Hypotheses/….md)

2. Hypotheses/ — distill 3–7 bets in IF/BY/WILL/BECAUSE form with, where it helps, a We're wrong if: line — the argument against it (Hypotheses/_template.md), each approved individually by the user. author: 'AI (Claude) — approved by <name>', source: 'cold-start founding brief, <date>', status: open. The riskiest assumption (Q8) comes first.
3. Research backlog.md — the open questions in "how would we know if this is wrong?" form.
4. Competitors/ — stubs per the needs_research convention.
5. Verified Evidence/ (optional, only if the user wants desk research now) — market/problem desk research per /researcher rules, with the lookback window offered as 1/2/3 years; pitch-deck stats confirmed at their original sources land here with real citations.

What this flow never produces: Signals/, Personas/, or any bump to sample stats. If the user wants a proto-archetype to think with, hand over to /persona-workshop Level 1 (its archetype step is designed for exactly this) — and be plain that a persona stays out of reach until the first real interview lands.

Closing the session — the route out of zero

End with the concrete path, in order:

1. Desk research round (if not done in step 5) → first real Evidence/.
2. /interview-guide — it already pulls open hypotheses as things to verify; the founding brief just fed it.
3. 3–5 real interviews with people from the (assumed) target group → drop transcripts in Inbox/ → /extract-findings → the first Signals/ and, at last, a persona grounded in something.

Offer a local commit ("Cold start: founding brief, N hypotheses, backlog seeded"). Never push unasked.

Rules

• Founder ≠ user. Their statements about users are always L1 — never Signals, never Evidence, never counted into confidence or sample stats.
• Never present a hypothesis as a finding — in this flow or any later one.
• No invented specifics. If the founder doesn't know (e.g. market size), the answer is a backlog entry, not a guess.
• Pitch-deck numbers are claims until the original source is verified; the verification, not the deck, is what Evidence/ cites.
• One assumption per Hypothesis file — same granularity rule as Signals.
• Demo files stay out — the founding brief concerns the user's real project only.