find-product-market-fit
The product-market-fit loop that drafts `BUSINESS_PLAN.md` and interrogates its weakest assumption until the 8-point readiness gate passes, called by ceh-business-plan:develop-business-plan.
SKILL.mdplugins/standalone/ceh-business-plan/skills/find-product-market-fit/SKILL.md
GitHubFind product-market fit
Produce a validated business plan — one whose product-market fit has been pressure-tested,
not asserted. The deliverable is a single living BUSINESS_PLAN.md (schema and PMF gate in
${CLAUDE_PLUGIN_ROOT}/references/business-plan-schema.md), revised in place across an interview loop until it passes
the 8-point PMF readiness gate and the user confirms.
The reason this skill exists: a business plan written in one pass is a wish list. Real product-market fit is found by repeatedly attacking the plan's weakest assumption — the one that, if wrong, kills the business — and revising until the load-bearing claims are evidenced or have a cheap test attached. This skill is that loop, made disciplined.
The core stance: draft proactively, then interrogate. Infer a v0 from what already exists and interview against it. A concrete draft gives the user something to react to, and correction is faster and sharper than composition.
The Loop
Phase 0 Intake → find/read app plans or provided material
Phase 1 Draft v0 → write BUSINESS_PLAN.md from what's known; score the PMF gate
Phase 2 Interview → attack the lowest-scoring, highest-leverage gate criterion (one question)
Phase 3 Revise → fold the answer in; re-tag confidence; re-score the gate
↑___________ repeat Phase 2–3 until gate = 8/8 AND user confirms
Phase 4 Validate → flip status to validated; hand off next steps
Each loop closes one gap. The gate score chooses the next question, so the interview is never a fixed questionnaire.
Phase 0 — Intake (read before asking)
Before any question, gather everything that already constrains the plan.
- Build plans in the repo — the
ceh-build-planningplans inproduct/plans/*.md(or the plans folder the repo'sCLAUDE.mdnames). If found, read them. Their Goal (what it does, who it's for, the key flow), Scope, and Design carry your product, target-user hint, and solution detail. A plan whose Goal points to aBUSINESS_PLAN.mdwas written from this plan: read it for its Scope and Design only. Record their paths inderived_from. Built plans stay in the folder withstatus: built: read them too, andproduct/ARCHITECTURE.mdOverview and Key Decisions when it exists. - Anything the user provided — a PRD, pitch deck, spec, README, landing page, a pasted description. Read it directly. A GitHub URL: fetch the README.
- Other plan-shaped files —
PRD.md,BRIEF.md,docs/product notes. Use them as input, not gospel.
Extract without asking: what the product does, who it seems to be for, the core flow, the stage
(idea / prototype / launched / revenue), and any pricing or competitor hints already written down.
If BUSINESS_PLAN.md already exists, you are continuing the loop — read it, re-score the gate,
resume at the lowest-scoring criterion.
If nothing exists (no plans, nothing provided): skip the draft-from-source path and open with a single grounding question — "In one or two sentences: what's the product, and who is it for?" — then draft v0 from that answer. Still draft before interrogating.
Phase 1 — Draft v0 proactively
Write a complete BUSINESS_PLAN.md per the schema — all 13 sections, even the thin ones. Fill
what the source material supports; for everything else, write your best-guess hypothesis and tag
it [assumption], not "TBD". A wrong, specific assumption is useful — it gives the interview a
target. A blank is not.
Tag every load-bearing claim: [evidence], [assumption], or [hypothesis-to-test]. Then score
the PMF gate (N/8) in frontmatter and identify the lowest-scoring, highest-leverage criterion
— that is where the interview starts.
Show the user the draft (or, for a long plan, the executive summary + the gate scorecard with which criteria are open). Then begin the interview. Do not wait for permission to draft.
Phase 2 — The PMF Interview
Rigorous means adversarial toward assumptions, not toward the user. You are hunting the claim most likely to be wrong and most expensive if it is. The gate scorecard picks the target; you ask the one question that would most move it.
Rules
- One question per turn. Never dump a list. Ask the single highest-leverage thing you don't yet know. Wait for the answer.
- The gate drives the order. Attack the lowest-scoring criterion first; within it, the claim whose failure is most fatal. Re-pick after every answer — the target moves.
- Offer a hypothesis to react to, not an open void. Instead of "What's your pricing?" ask "I've assumed $49/seat/month because that undercuts [competitor]'s $79 — does that match what you think they'll pay, or is the real anchor their current spend on [alternative]?" Correcting a concrete guess is faster and sharper than composing from scratch.
- Push past abstractions to a real instance. "SMBs want this" → "Name one. Who specifically told you, or who's a real candidate, and what do they do today instead?" The first sign of a weak plan is a population where there should be a person.
- Separate evidence from belief, out loud. When an answer is a hope, say so and re-tag it
[assumption]; when it's a real signal (a quote, a transaction, a comparable spend), tag it[evidence]; when it's a belief with a cheap test, make the test concrete and tag it[hypothesis-to-test]. The tags are the honesty mechanism. - Capture the customer's words verbatim. A real quote ("I'd pay anything to stop doing this in Excel") is the strongest asset a plan has — store it word-for-word in §02 or §10.
- Steelman the alternative. For differentiation, force the question: "Why doesn't the customer just keep doing nothing / using the spreadsheet?" Inertia is the default competitor and beats most products.
- Know when to stop asking on a point. Once a criterion is evidenced or has a named test, move on — don't gold-plate a passing criterion while others sit at zero.
The criteria as an agenda (see the 8-point gate in the schema)
| Gate criterion | The question it demands an answer to |
|---|---|
| 1 Problem real & acute | Who feels this, how often, how badly — and what does it cost them now? |
| 2 Beachhead narrow & reachable | One nameable segment — where are the first 10 customers, physically? |
| 3 Differentiation defensible | Why you over the real alternative (incl. doing nothing), and why can't it be copied? |
| 4 Willingness to pay | What signal says they'll pay this price — a quote, a pre-order, a comparable spend? |
| 5 Market math bottom-up | SOM = customers × price with stated assumptions — does the arithmetic clear the bar? |
| 6 Unit economics can work | Is there a believable path to LTV > CAC with a sane payback? |
| 7 First channel credible | One channel where they already are, with a first-10 plan? |
| 8 Riskiest assumption has a test | What single wrong assumption kills this — and the cheapest test for it? |
Phase 3 — Revise
After each answer (or a small batch), fold it into BUSINESS_PLAN.md:
- Update the affected sections; re-tag the confidence on every claim the answer touched.
- Regenerate §01 Executive Summary and §12 Risks ranking — they shift as the plan firms up. Leave every specialist subsection and tagged line as it is (see "Specialist subsections" in the schema): the premortem's kill criteria in §12 are fixed in advance on purpose.
- Re-score
pmf_gate: N/8in frontmatter and bumpupdated. - Tell the user, in one or two lines, what just moved (e.g. "Criterion 4 went from assumption to evidence — you have two LOIs at $49. Lowest now is criterion 5: the market math is still top-down. Next question is about that.").
Then return to Phase 2 on the new lowest-scoring criterion. This is the loop the goal demands: draft → interview → revise, repeated until satisfiable.
When a criterion can't be evidenced yet
At idea stage, most criteria will not have evidence — that is expected. Do not stall the loop
waiting for data that requires building or selling first. Instead, convert the bare [assumption]
into a [hypothesis-to-test]: name the cheapest experiment that would settle it (a landing-page
smoke test, 5 customer interviews, a concierge MVP, a pre-sale) and schedule it in §13. A
criterion with a real, cheap, scheduled test passes the gate. Hand-waving does not.
Phase 4 — Validate & Hand Off
When all 8 criteria are met (evidenced or test-attached) and the user confirms they are satisfied:
- Flip frontmatter
status: validated, setpmf_gate: 8/8, stampupdated. - Give a one-paragraph verdict: the sharpest version of who pays, for what, why you, and the one experiment that most de-risks the whole thing next.
- Point onward: if the validation surfaced product changes, the build plan
(
ceh-build-planning) should absorb them in a new plan; the §13 milestones become the build/validation backlog. - Name the plan's softest remaining area (edge, numbers, first customers, risks, or the next 90
days). The gate proves a fit exists, not that the business is sound, and
ceh-business-plan:develop-business-planroutes that area to the specialist that owns it.
Do not flip to validated to end the loop early. If the user wants to stop before the gate
passes, leave status: draft, record the open criteria honestly, and say plainly which
assumptions remain untested — an honest draft beats a plan that claims a fit it hasn't found.
Stop Conditions
- The idea fails a gate criterion with no path to a test — e.g. the problem isn't real, or no reachable customer exists. Say so directly. A plan's most valuable output can be "don't build this, here's why." Don't manufacture a fit.
- The product is fundamentally different from the app plan it derives from — flag the mismatch; the business plan may be pointing at a pivot the app plan hasn't caught up to.
Edge Cases
App plans exist but are sparse (skeleton only): draft from the §01 Concept and treat §02–§13
as mostly [assumption] — the interview has more to do, which is fine.
User has real traction already (launched/revenue): start the gate from evidence, not guesses. Many criteria may pass on intake; the loop focuses on the few still soft (often §07 channel scaling or §06 defensibility).
User resists a question / answers a different one: keep the answer — it's real material — re-tag what it informs, and re-ask the original once at most, rephrased. Don't interrogate.
"Just write me a business plan, no questions": draft v0, score the gate, and hand it over with the open criteria flagged as the assumptions they're betting on — but say once that the plan is unvalidated until those are tested.
No product idea at all, just "I want to start something": the problem comes before the product. Interview for an acute problem the user has standing or insight to attack, then draft.
No human to answer (a headless run, or called by another agent): take the "no questions" path
above. Draft v0, score the gate, leave status: draft, and end with the open criteria as the
questions a person must answer, in gate order, each with the assumption the draft used. When nothing
exists to draft from, stop and report that the plan needs a product and a customer from a person.