Agent Booster Pack for Pi

ABP skills packaged for Pi, the terminal coding agent. Helps Pi produce code that is well-organized, low in complexity and side effects, and is secure and performant.

The pack covers data modeling, proof obligations, code review, security, debugging, refactoring, and more. It is one of three distribution surfaces for the same skill set; the canonical source lives at agents/.agents/skills/ in the parent repository, and skills/ is built from that source at npm-pack time and is not committed.

Install

From npm (recommended)

pi install npm:pi-agent-booster-pack

Then in Pi:

/reload

From a local checkout

For development against the live source. Run the build first so the skills exist in pi/skills/:

cd /path/to/agent-booster-pack/pi
npm run build
pi install /path/to/agent-booster-pack/pi

Then in Pi:

/reload

Local development symlink

If you are iterating on skills and want a single command for build + install:

cd /path/to/agent-booster-pack/pi
npm run install-ext

This builds skills/ and symlinks the working tree into ~/.pi/agent/extensions/pi-agent-booster-pack. Re-run npm run build and /reload in Pi after editing skills.

Note: pi install git: does not currently support installing a subdirectory of a repository, so pointing at the parent agent-booster-pack git URL will not work. Use the npm install (above) or a local checkout.

What's included

20 skills, grouped by the kind of engineering pressure they apply.

Entry point

• workflow — choose the right ABP skills for the task, name what is being coupled, keep the work scoped, and connect completion claims to proof.

Foundational design

• data-first — domain data, fields, states, transitions, effects.
• architecture — module boundaries, domain locality, DDD tactical patterns.
• proof — proof obligations and behavior-focused tests.

Correctness and change

• code-review — risk-focused review of diffs, branches, PRs.
• debugging — root-cause investigation.
• refactoring — behavior-preserving structure changes.
• error-handling — error types, propagation, retries, recovery.

Safety gates

• security — auth, secrets, crypto, input validation, trust boundaries.
• database — schemas, migrations, queries, transactions, deletion semantics.
• release — version bumps, CHANGELOG, rollouts, rollbacks, feature flags.

Production quality

• observability — logs, metrics, traces, dashboards, SLOs, alerts.
• async-systems — async work, queues, streams, ordering, backpressure.
• performance — latency, throughput, caching, allocation, stampede prevention.

Public/user surfaces

• api — HTTP APIs, OpenAPI, idempotency, pagination, webhooks.
• documentation — READMEs, ADRs, runbooks, reference docs.
• ui-design — pages, components, interaction flows, responsive layout.
• accessibility — WCAG, semantic HTML, ARIA, keyboard navigation.

Project workflow

• git-workflow — branch hygiene, commit splits, rebases, history recovery.
• scaffolding — new projects, baseline tooling, CI defaults.

How it works

Pi reads pi.skills from this package's package.json and walks each subdirectory under skills/. Each subdirectory is one skill, with a SKILL.md that has YAML frontmatter (name, description) and a body Pi loads when the skill description matches a task. No code runs.

Build pipeline

skills/ is a build artifact, not a checked-in mirror.

• scripts/build-skills.mjs walks ../agents/.agents/skills/ and copies each skill subdirectory into ./skills/.
• npm run build runs that script directly.
• npm pack and npm publish run it via the prepack lifecycle hook, so the published tarball always reflects the canonical source at the moment of publish.

This means there is exactly one place to author skills (agents/.agents/skills/ in the parent repo), and no chance of pi/skills/ drifting from canon.

Publishing

From inside pi/:

npm publish

prepack runs npm run build automatically, so the tarball includes the freshly-built skills/ directory.

License

MIT — see LICENSE.