reSkill

Turns an idea into a finished build prompt for a Claude Code / Codex skill.

Connect an agent (MCP)

Agents can drive the whole flow over MCP. Clerk is the OAuth server, so the agent signs the user in through a normal consent screen and acts only on that user's skills.

claude mcp add --transport http reskill https://<your-host>/mcp

Tools: list_skills, create_skill, get_skill, review_summary, answer_question, start_build, get_build. The flow is create, confirm the summary, answer the questions, start the build, then poll get_build for the prompt.

To save round trips, get_skill takes batch (up to 5) to return several open questions, and answer_question takes answers to record them in one call (one model call for the batch).

Dangerous mode: create_skill with dangerous: true is for hands-off runs. The agent answers the intake questions itself and keeps calling start_build with improve=true (up to 3 times) until the score target is met, never asking the user. Only use it when the user explicitly asks; the answers are the agent's guesses, not the user's decisions.

One-time Clerk setup

1. Clerk Dashboard, OAuth applications, Settings: turn on Publish CIMD support, and Publish DCR support for clients that lack CIMD.
2. Set default scopes for dynamic clients to profile email.
3. Set CLERK_PUBLISHABLE_KEY on the Worker (wrangler secret put CLERK_PUBLISHABLE_KEY, or .dev.vars locally). It is the same pk_ key as VITE_CLERK_PUBLISHABLE_KEY.
4. Apply the new migrations (including 0006_autonomous.sql): npm run db:migrate:remote.

Limits per user: 10 new skills per 24 hours, and per 7 days 100 planning steps (model-backed calls), 20 builds (WEEKLY_RUN_LIMIT in wrangler.jsonc) and 50,000 model tokens (WEEKLY_TOKEN_LIMIT, counted in token_usage). Every limit message tells the user to email hi@partavocado.cc for an increase.