| name | mvp |
| description | Guides the founder through designing and optionally building the simplest MVP or prototype that validates their current hypotheses. Use when the founder wants to build something to test assumptions, discusses what to build next, wants to interpret results from a live MVP, or is deciding whether the current approach is still right. Also use when a founder proposes something to build — the skill will check whether the proposed form is the simplest thing that generates honest signal. Also use when the founder wants cheap behavioral demand signal — a fake door (a not-yet-built feature or tier that logs real intent) or a small paid-traffic experiment to reach people they cannot reach by hand. |
MVP / Prototype
Help the founder figure out the simplest thing worth building to validate their remaining assumptions — and optionally scaffold and deploy it.
The central job of this skill is to be a principled counterweight to over-engineering. Most founders want to build more than they need to test what they don't yet know. This skill reads what's been validated, identifies the riskiest untested assumptions, and argues for the form of MVP that generates honest signal with the least build effort.
Two modes:
- Design conversation — structured dialogue that produces
startup/mvp-plan.md: what to build, why this form, which hypotheses it tests, and what success looks like
- Scaffold and deploy — optional Layer 2 reference that writes code to the project root and deploys using Vercel MCP, Supabase MCP, and the v0 Platform API
Before you start
Read startup/core.md and scan startup/hypotheses/ to understand what's been established and what's still untested. Check startup/interviews/ and startup/surveys/ for evidence gathered so far. No directory scaffolding is needed — startup/ is created during project initialization.
When no startup/mvp-plan.md exists
Load the reference file that runs the design conversation:
references/initial-mvp-design.md
The reference file's instructions take over from this point.
When startup/mvp-plan.md exists
Read it for context. Infer intent from the conversation — don't ask "what do you want to do?"
If the founder is discussing results or what they're seeing:
Handle inline. Read the plan to understand what was built, what hypotheses were being tested, and what the success criteria were. Also check startup/interviews/ and startup/surveys/ for any evidence collected since the MVP launched — this context informs the assessment. Ask what they're seeing — numbers, anecdotes, surprises. Compare against the success criteria and give a frank read:
- Confirmed — signal clearly supports the hypothesis; route updates through the
hypotheses skill
- Contradicted — signal clearly runs against it; route updates through the
hypotheses skill
- Inconclusive — make the distinction explicit: "the hypothesis is probably wrong" is different from "the experiment didn't reach the right audience or ran too short." The first warrants invalidating the hypothesis; the second warrants redesigning the experiment, not changing hypothesis state.
Update the ## Experiments Log in mvp-plan.md with what was learned (dated entry). If the plan needs to evolve, propose changes and get confirmation before writing back.
If the founder wants to iterate or pivot the experiment:
Discuss what's changed. Propose what the next experiment should look like. Before overwriting the plan, move the current success criteria and outcome into the ## Experiments Log as a completed entry. Then update ## What We're Building, ## Why This Form, ## Hypotheses Being Tested, ## Success Criteria, and ## Distribution Plan with the new experiment. Propose the full updated content before writing. Get confirmation.
If the founder wants to scaffold and deploy:
If a founder proposes building something without a prior design conversation:
Read the existing hypotheses. Brief honest check (2–3 sentences): is the proposed form the simplest thing that would test the riskiest untested assumptions? Share the assessment before proceeding — not a gate, just an informed nudge.
If the founder wants to archive:
Archiving marks this MVP track as closed — the plan remains for reference but is no longer the active experiment. Read the file. Set status: archived, last_updated: today. Add a final log entry summarising the experiment outcome. Propose changes, get confirmation, write back.
Fake-door and paid-traffic knowledge
Two validation techniques are available as knowledge the skill draws on when the situation fits — not workflows to run through. They are contextual and often unique per case. They compose (the classic smoke test is a fake door plus paid traffic) but neither implies the other — use both, one, or neither.
- Fake door — a not-yet-built feature, tier, or button that logs real intent; a lighter landing-page form. Load
references/fake-door.md when a fake door is a live option (the founder is choosing or proposing a form, or you spot a fit). It carries the judgment for when a fake door still earns its place now that building is cheap, plus the honest-reveal rules.
- Paid traffic — buying a small slice of cold audience when the founder can't reach enough of the right people by hand. Load
references/paid-traffic.md when reach is the bottleneck on an experiment. It applies to any deployable form, fake-doored or real, and covers contextual channel choice, small-budget segment-splitting, live benchmark research, and the honest automation ceiling.
Either technique's results are recorded in the existing startup/mvp-plan.md (## Hypotheses Being Tested backlinks, ## Distribution Plan, ## Experiments Log) — no separate artifact.
After saving startup/mvp-plan.md
Briefly confirm: "Saved to startup/mvp-plan.md."
Then close with the single concrete next move, so the founder is never left wondering what to do. The MVP is a learning instrument, not the finish line — say so plainly, and name the move that gets signal flowing:
status: ready and deployable form → "Ready to scaffold and deploy — just say the word."status: ready and a manual form (Wizard of Oz, concierge, manual outreach) → point at the ## Distribution Plan: "Your next move is to run it — get it in front of {the people and count in the distribution plan} and watch for {the success criteria}."status: live → "The plan only generates signal once it reaches the right people — execute the ## Distribution Plan now, then watch {the success criteria}."
Then tell the founder exactly how to re-enter the loop:
- When signal arrives, come back and say "how's the MVP doing?" — we'll read it against the success criteria and decide together whether to persevere, iterate the experiment, or pivot.
- Or run
/whats-next any time to re-plan around what the MVP is teaching you.
Optionally, one line: running interviews or surveys in parallel often produces richer validation than the MVP alone — the MVP tells you if they use it, conversations tell you why.
