mizchi/skills

A collection of agent skills maintained by @mizchi, distributed via APM (Agent Package Manager).

Each directory is a standalone skill following the agentskills.io open standard.

Install

Install an individual skill (global / user scope):

apm install -g mizchi/skills//

Or add to a project’s apm.yml:

dependencies:
  apm:
    - mizchi/skills//

Pin to a tag:

apm install -g mizchi/skills//#v0.1.0

These follow the open Agent Skills format, so any compatible agent (Claude Code, Codex, opencode, Gemini CLI, Cursor, …) can use them without APM. Each skill’s README shows a npx skills add command (the skills CLI, installs into every detected agent) and a raw npx degit folder copy as alternatives:

npx skills add mizchi/skills --skill 

Every skill directory carries its own README.md with copy-paste install examples, generated from the skill’s SKILL.md frontmatter. Regenerate them all after editing any frontmatter:

pkf run gen:readme     # write / sync every skill README
pkf run check:readme   # CI: verify they are in sync

Skills

Frontend

A toolkit for frontend consulting engagements. Install skills from mizchi/skills/frontend/.

Perspective sub-skills (invoked by frontend-review-weekly): frontend/review-perspectives/{frontend-expert,react-expert,performance-expert,security-expert,frontend-ops-expert}.

Node.js

AWS

Cloudflare

SQL

Languages

Testing / Browser

AI / VLM

Formal Methods

Tooling / Infra

Kubernetes

DevOps

Process / Meta

Tools

waxa — skill evaluation CLI

Lives at tools/waxa/ and is published as @mizchi/waxa. Used by the waxa-eval skill, and forms the adoption gate of skill-finder.

Run via npx (no global install needed):

npx @mizchi/waxa 

npx caches the package after the first run, so subsequent invocations are fast. Pin a version with npx @mizchi/[email protected] when reproducibility matters (do not pin 0.1.0 — that release shipped a broken shebang and crashes before argv parsing).

Alternatives:

# install globally (when calling waxa frequently)
npm i -g @mizchi/waxa

# run from source (requires Deno 2+)
git clone https://github.com/mizchi/skills.git
cd skills/tools/waxa
deno task run -- path/to/eval.yaml

Requirements: claude CLI on PATH and authenticated (OAuth login or ANTHROPIC_API_KEY).

Quick reference:

npx @mizchi/waxa init [--skill ] [--force]                   # scaffold /evals/ (skill-local)
npx @mizchi/waxa /evals/eval.yaml                           # single run
npx @mizchi/waxa /evals/eval.yaml --baseline                # with_skill vs without_skill + delta
npx @mizchi/waxa iterate /evals/eval.yaml [--max N]         # iteration loop with ledger
npx @mizchi/waxa compare /evals/eval.yaml --models ,            # multi-model comparison
npx @mizchi/waxa variant /evals/eval.yaml --base  --candidate   # A/B

From 0.2.0 evals live at /evals/ (skill-local layout, agentskills.io-aligned). Workspace outputs go to /results//iteration-N/. The npm package bundles references/empirical-prompt-tuning.md (the methodology document), so iter / convergence semantics ship with the CLI.

Full reference: tools/waxa/README.md.

Recommended starter set

If you are setting up a new project that will produce or consume skills, this combination covers selection, discovery, and iteration:

# apm.yml
targets:
  - claude
dependencies:
  apm:
    - mizchi/skills/tooling/apm-usage          # APM manifest reference
    - mizchi/skills/meta/skill-selector        # Phase 1: pick from curated catalog
    - mizchi/skills/meta/skill-finder          # Phase 2: cross-source discovery + waxa eval gate
    - mizchi/skills/meta/waxa-eval             # waxa CLI operating manual
    - mizchi/skills/meta/empirical-prompt-tuning  # methodology / Iter 0 / [critical]-tag checklist

Then:

apm install                                # installs the skills under .claude/skills/
npx @mizchi/waxa --help                    # cache the waxa CLI on first call
# (or `npm i -g @mizchi/waxa` if you call waxa often enough that startup matters)

The skill-selection / discovery / evaluation flow then proceeds as: skill-selector (catalog pre-flight) → if no fit, skill-finder (Tier 1-4 sweep) → adoption gated by waxa-eval running waxa iterate against a evals// directory you author. empirical-prompt-tuning covers what waxa cannot reach (the Iter 0 description / body consistency check, tool_uses-based skill self-containment diagnosis, and the [critical]-tagged requirements checklist that gives binary success judgment).

Language

All SKILL.md files are written in English. Skills that were originally written in Japanese preserve the original as SKILL-ja.md alongside the English version. mizchi-blog-style is the single exception and stays Japanese-only because the skill itself is about Japanese blog writing style.

Upstream skills (not in this repo)

These are maintained in their own repositories and installed separately:

• mizchi/flaker — flaker-setup
• mizchi/tui.mbt — tuimbt-practice
• moonbitlang/moonbit-agent-guide — moonbit-agent-guide, moonbit-refactoring, moonbit-c-binding
• ast-grep/agent-skill — ast-grep

License

Each skill may carry its own license (see LICENSE.txt inside the skill directory). Skills without an explicit license default to MIT at the repository owner’s discretion.