kirin-pi

Bryan's skill pack for coding agents. One git-installable repo with filterable skills, plus a uv bash guard extension.

Works with Pi, Claude Code, and Codex. All skills follow the Agent Skills standard (SKILL.md + YAML frontmatter).

Contents

Practices — software development methodology

The chain: grill-me → writing-plans → execute-plan → tdd → verification → code-review → commit, with to-prd, improve-codebase-architecture, diagnose as branch-offs.

Skill	Path	Purpose
grill-me	skills/practices/grill-me	Required before any creative work. Two modes: exploration (new idea → design doc) or grilling (existing plan → CONTEXT.md/ADR tightening). HARD-GATES implementation.
to-prd	skills/practices/to-prd	Synthesize conversation into a PRD with user stories and implementation/testing decisions.
writing-plans	skills/practices/writing-plans	Write bite-sized implementation plans (TDD, exact file paths, ready for execution).
execute-plan	skills/practices/execute-plan	Execute a written plan. Two modes — subagent-driven (default) dispatches fresh subagent per task with two-stage review; inline (fallback) runs in current session. Auto-detects from subagent availability.
tdd	skills/practices/tdd	Red-green-refactor with Iron Law, integration-style tests through public interfaces, and mandatory verify-fail/verify-pass.
diagnose	skills/practices/diagnose	Disciplined 6-phase debugging loop with Iron Law and 3-fix architectural escalation.
verification-before-completion	skills/practices/verification-before-completion	Evidence before claims. Never assert work is done without fresh verification output.
code-review	skills/practices/code-review	Review a diff/branch/PR across quality, security, correctness lenses.
improve-codebase-architecture	skills/practices/improve-codebase-architecture	Find deepening opportunities; architecture review with HTML report.
commit	skills/practices/commit	Create Conventional Commits from working-tree changes.
changelog	skills/practices/changelog	Regenerate CHANGELOG.md [Unreleased] from commits and pending changes.

Meta — agent infrastructure

Skill	Path	Purpose
handoff	skills/meta/handoff	Compact the current session for a fresh agent to continue later.
reflect	skills/meta/reflect	End-of-session retrospective; writes retro file and proposed diffs.
web-search	skills/meta/web-search	Current web research with cited source URLs.
web-read	skills/meta/web-read	Fetch URL/local document and return clean markdown.
write-a-skill	skills/meta/write-a-skill	Create new skills with clear triggers, structure, and resources.
dispatching-parallel-agents	skills/meta/dispatching-parallel-agents	Dispatch one agent per independent problem domain to work concurrently.
zoom-out	skills/meta/zoom-out	Get a higher-level map of modules and callers. Tiny, manually invoked.

One-liners for your context file

Some behaviors are too small to be skills — put these in your project's CLAUDE.md (Claude Code), AGENTS.md (Codex / Pi), or .pi/CLAUDE.md (Pi):

- When I say "caveman", switch to ultra-compressed terse mode: no greetings, no prose, just shell-style output.

Domain skill groups

Group	Skill	Path	Purpose
Frontend	frontend-design	skills/frontend/frontend-design	Visual explanations or production frontend UI with aesthetic direction.
Frontend	prototype	skills/frontend/prototype	Throwaway code to answer a design question — terminal TUI for logic/state, variant pages for UI design.
Languages	rust	skills/languages/rust	Rust project workflow: cargo/rustup, format, clippy, tests.
Languages	uv	skills/languages/uv	Use uv instead of pip/python/venv/poetry.
Travel	planning	skills/travel/planning	End-to-end trip planning inside the kirin vault.
Travel	destination-research	skills/travel/destination-research	Research destination and write/update evergreen guide.
Travel	trip-prep	skills/travel/trip-prep	Packing, logistics, reservations, loyalty/points tactics.
Travel	vlog-prep	skills/travel/vlog-prep	Shot list, b-roll, golden hour, audio notes, narrative arc.
Travel	post-trip	skills/travel/post-trip	Post-trip retrospective and destination-guide updates.

Extension (Pi-specific)

Extension	Path	Purpose
uv.ts	extensions/uv.ts	Blocks pip/poetry/python -m venv etc. at bash-tool time and pushes the agent toward uv.

The travel skills assume the kirin vault Travel folder conventions. They are domain-specific, not generic project tools.

Install

Pi

// ~/.pi/agent/settings.json
{
  "packages": [
    {
      "source": "git:github.com/bryan-tu/kirin-pi",
      "extensions": ["extensions/uv.ts"],
      "skills": [
        "skills/practices",
        "skills/meta"
      ],
      "prompts": [],
      "themes": []
    }
  ]
}

For local development, use the local source path instead:

"source": "/Users/bryan/Projects/kirin-pi"

Add optional skills by path, for example:

"skills/frontend/frontend-design",
"skills/languages/rust",
"skills/languages/uv",
"skills/travel/planning"

Omit the skills filter to load every skill under skills/ recursively.

Claude Code

Clone the repo and add the skills path to your settings:

// ~/.claude/settings.json
{
  "skills": [
    "/path/to/kirin-pi/skills"
  ]
}

Or for project-level use:

// .claude/settings.json
{
  "skills": [
    "../kirin-pi/skills"
  ]
}

Codex

Clone the repo and symlink or configure the skills directory:

# User-level
ln -s /path/to/kirin-pi/skills ~/.codex/skills/kirin-pi

# Or project-level
ln -s /path/to/kirin-pi/skills .agents/skills/kirin-pi

Layout

kirin-pi/
├── package.json
├── extensions/
│   └── uv.ts              # Pi-specific
├── intercepted-commands/   # Pi-specific
│   ├── pip, pip3, poetry, python, python3
└── skills/                 # Cross-platform (Agent Skills standard)
    ├── practices/          # Software development methodology
    │   ├── grill-me/
    │   ├── to-prd/
    │   ├── writing-plans/
    │   ├── execute-plan/
    │   ├── tdd/
    │   ├── diagnose/
    │   ├── verification-before-completion/
    │   ├── code-review/
    │   ├── improve-codebase-architecture/
    │   ├── commit/
    │   └── changelog/
    ├── meta/               # Agent infrastructure
    │   ├── handoff/
    │   ├── reflect/
    │   ├── web-search/
    │   ├── web-read/
    │   ├── write-a-skill/
    │   ├── dispatching-parallel-agents/
    │   └── zoom-out/
    ├── frontend/
    │   ├── frontend-design/
    │   └── prototype/
    ├── languages/
    │   ├── rust/
    │   └── uv/
    └── travel/
        ├── planning/
        ├── destination-research/
        ├── trip-prep/
        ├── vlog-prep/
        └── post-trip/

Credits

This repo grew from studying and adapting excellent agent-tooling repos:

• obra/superpowers — methodology and the vendored writing-plans, verification-before-completion, dispatching-parallel-agents skills, plus the base for the merged execute-plan (combines executing-plans and subagent-driven-development). Also merged into tdd (Iron Law, verify-fail/verify-pass discipline), diagnose (3-fix escalation, multi-component instrumentation), grill-me (HARD-GATE, exploration mode), and write-a-skill (TDD-for-skills, description = when-not-what).
• mattpocock/skills — skill-library ergonomics, workflow framing, and the vendored to-prd, improve-codebase-architecture, prototype, zoom-out, handoff skills. Plus the base for the merged grill-me (CONTEXT.md/ADR tightening), tdd (deep modules, integration-style tests), diagnose (feedback-loop discipline, 6 phases), and write-a-skill (lean structure).
• juicesharp/rpiv-mono — package/filter thinking, Pi workflow design, and the multi-skill repo structure that inspired code-review.
• mitsuhiko/agent-stuff — uv skill, uv.ts guard extension, and intercepted command shims vendored here.

License

MIT