Documentation
Product Autopilot is a Claude Code skill plus a self-hosted dashboard. The skill closes the product loop — instrument, observe, analyze, outreach, decide, build — and writes its work back as markdown in your repo. These docs are the short version while the full reference is drafted.
Three steps to a live loop.
- step 1
Install the skill
Add the skill to your Claude Code config. It registers PostToolUse hooks and the six sub-skills the agent can call on demand.
claude skill add @product-autopilot/skill - step 2
Run the web app
Boot the dashboard locally or deploy it behind your VPN. It provides the ingest endpoint, the query API, and the UI.
pnpm --filter web dev - step 3
Instrument your product
Add the SDK to your app. Autopilot drafts the first events and mounts the feedback widget automatically when the skill sees new code.
npm i @product-autopilot/sdk
Four nodes, one data model.
The SDK is the only thing that touches your product. The web app owns the store. The skill is how the agent reads and writes back.
Emits events, logs, errors, and feedback through a typed client.
~6kb gzipped. Batched, typed, safe to ship on the edge.
Postgres-backed ingest, query endpoints, realtime stream over SSE.
Hooks + sub-skills. Reads from the web app, writes agent activity back.
See the full diagram on the landing page.
Six sub-skills the agent reaches for.
Each sub-skill is a small, typed capability. The agent picks one based on the task at hand; humans read the same markdown outputs.
- instrumenterDrafts typed events and mounts feedback widgets next to new code.
- observerQueries the ingest store for events, errors, sessions, and feedback.
- analystSynthesizes signals into ranked insights with linked evidence.
- outreachDrafts interviews, surveys, and re-engagement notes against insight-derived segments.
- plannerTurns insights and outreach results into backlog items carrying their receipts.
- reporterProduces weekly digests and traceability reports for reviewers.