pi-auto-skill

After you finish a task in Pi, this package asks whether it is a procedure worth keeping as a project skill — and if you agree, the agent writes one, following the writing-great-skills reference, and you review it before it is saved to the repo.

Instead of explaining the same deploy, migration, or release routine every time, the next session finds a skill for it.

Install

pi install npm:@fahrizkyputra/pi-auto-skill

How it works

1. You work as usual. When the agent finishes a run, the package checks a minimum bar: at least 5 tool calls, or at least one file changed. Plain questions and short chats are never judged.
2. If the bar is met, the active model is asked once (a small call) whether the task is a procedure that will be repeated in this project. It also sees the project's existing skills and the names you declined before.
3. If yes, a dialog asks: Create skill deploy-staging? — or Update skill deploy-production? when an existing skill covers the procedure but the task revealed new steps.
4. On Yes, the agent in your session writes the skill from the task it just finished, following writing-great-skills: a reusable procedure, with values that change between runs (branches, IDs, versions) turned into named inputs.
5. The draft opens in an editor. Edit it, then save to keep it or cancel to discard it. Nothing is written to disk before this point.
6. On No, that name is never suggested again in this project.
7. After you save, the agent restates your original task — what is done, what is not done or needs you, and the next step (at most 8 lines, in the language you wrote in) — and the package puts a fixed block right under it:

   🧩 Skill "deploy-staging" created
      .agents/skills/deploy-staging/SKILL.md
      Run /reload to use it.
   

   Writing the skill pushes your task far up the chat, so without this you would have to scroll back to find what was left for you to do. The recap uses the original request, the files changed, and the agent's own final answer from the task, which the package kept; it does not describe the skill. If you type something else before the recap ends, the block is not shown late.

The skill lands in the repo's skills folder: .agents/skills/ if it exists, else .pi/skills/ if it exists, else .agents/skills/ (the Agent Skills location other agents read too). Run /reload to load it.

Commands

Per-project state (on/off and declined names) is kept in <repo>/.pi/auto-skill.json. Commit it to share declines with the team, or add it to .gitignore.

Requirements

The package does not ship a skill-writing guide of its own. It reads writing-great-skills from ~/.pi/agent/skills/writing-great-skills/ (or $PI_CODING_AGENT_DIR/skills/writing-great-skills/): SKILL.md is embedded in the agent's brief, and GLOSSARY.md is referenced by path. If it is not installed, nothing is judged and /auto-skill tells you why.

Behaviour details

• The agent hands the draft over through a propose_skill tool, never by writing the file itself. Drafts with invalid frontmatter (name format, missing or over-long description, empty body) go back to the agent to fix before you see them.
• Updating a skill moves the previous SKILL.md to the trash first (trash CLI), or keeps a .bak copy when the trash is unavailable.
• If you type a new message while the judgement is still running, the suggestion is dropped.
• Subagents and non-interactive runs (RPC without dialogs, print mode) are never judged, so no hidden model calls happen there.
• Version 1 writes one SKILL.md per skill; supporting reference files are not generated.

Cost

One small model call per task that meets the minimum bar, plus one agent turn when you accept a suggestion.

Development

npm install
npm test          # node:test, no Pi required
npm run typecheck
npm run test:e2e  # real Pi over RPC with a scripted model: judge → approve → propose_skill → edit → save; decline

License

MIT