validate

Run a full validation pass on an idea — sharpening, devil's advocate, market sizing, competitor mapping, customer discovery plan.

SKILL.md

GitHub

/validate

An opt-in rail, not a gate. Use /validate when a bet is expensive, slow, or hard to undo — paid acquisition, a big build, a pivot. For cheap, reversible ideas, just build; the ship is the validation.

Usage: /validate <idea or product name>

What this does

Runs the Idea-stage agents in sequence (or parallel where independent):

  1. problem-sharpener — turn the idea into a testable hypothesis
  2. devils-advocate — build the strongest case against
  3. market-researcher — TAM/SAM/SOM + trends
  4. competitor-mapper — landscape by tier
  5. customer-discovery — interview plan + outreach templates

Outputs go to products/<slug>/ (creating the folder if new).

What you do as the assistant when this is invoked

  1. Resolve <idea> to a product slug (existing or new).
  2. If new product: create products/<slug>/ with minimal structure, log in decisions/0001-validation-start.md.
  3. Run the 5 agents — fleet mode when subagent spawning is available (see factory/playbooks/mvp-stage/fleet-patterns.md):
    • Fan out blind: problem-sharpener, devils-advocate, market-researcher, competitor-mapper as 4 concurrent subagents, each getting the raw idea + its own SKILL.md only — none sees another's output (independent verdicts are the point; devils-advocate attacks the raw idea, which is the harder, more honest target). Spawn each on its model_tier.
    • Verify adversarially: before synthesis, spawn one refuter against the kill case ("is this actually fatal, or survivable?") and one against the strongest FOR evidence ("is this signal or founder hope?").
    • Synthesize: customer-discovery runs on the (surviving) sharpened hypothesis; then produce the snapshot, reporting where the blind agents agreed (high confidence) and disagreed (the open questions — never averaged away).
    • Serial fallback (no spawning): run in sequence as before — 1, then 2 on 1's output, 3+4, then 5 on 1+2. Same deliverables either way.
  4. After all done: produce a 1-page "validation snapshot" summarizing GO / PAUSE / KILL with reasoning — including the agreement/disagreement map when fleet mode ran.

Output location

  • products/<slug>/validation/snapshot.md — the summary
  • products/<slug>/validation/hypothesis.md — sharpened problem
  • products/<slug>/validation/devils-advocate.md — kill case
  • products/<slug>/validation/market.md — sizing + trends
  • products/<slug>/validation/competitors.md — landscape
  • products/<slug>/interviews/discovery-plan.md — the plan

When NOT to invoke

  • If the idea hasn't been articulated in writing yet — ask user to write 2 sentences first.
  • If user just wants a quick gut check — that's idea-generator or devils-advocate directly, not full validation.