Agent Skills

A collection of agent skills I use day-to-day across Claude Code, Codex, Cursor, and other agents that follow the open agent skills convention.

Use skills to install any single skill, or browse them all interactively:

# install one skill (replace <skill> with any name from the catalog below)
npx skills add sorafujitani/skills/<skill>

# pick from a list
npx skills add sorafujitani/skills

# install everything for Claude Code only
npx skills add sorafujitani/skills --skill '*' -a claude-code -g -y

Each skill ends up at ~/.claude/skills/<skill>/SKILL.md (or the equivalent directory for other agents).

Hands-on modes that hold the keyboard for you instead of writing code on your behalf.

• guided-code — The agent never writes code; it tells you what to write and where, phase by phase, until the task is done.

  npx skills add sorafujitani/skills/guided-code

• print-debug — One step, one observation. Place a println / console.log, run, read, then decide the next probe — both to chase a bug and to learn an unfamiliar codebase.

  npx skills add sorafujitani/skills/print-debug

Read-only modes for thinking through a change before any code moves.

• dry-coding — Plan-mode wrapper for PR- to epic-sized design. Parallel Explore/Plan sub-agents, schema contracts, and a design doc plus reviewable dry-coded implementation under .claude/plans/. No file edits.

  npx skills add sorafujitani/skills/dry-coding

• issue-analysis — Four-phase analysis of an OSS issue or feature request: independent parallel investigation, hypothesis/refutation cycles, evidence scoring, then a TDD-shaped implementation plan with cited sources.

  npx skills add sorafujitani/skills/issue-analysis

• pr-comment-plan — Pull the unresolved review comments on the current branch's PR, classify them (Critical / Important / Minor), and produce an ordered remediation plan.

  npx skills add sorafujitani/skills/pr-comment-plan

• pr-generator — Generate a PR from the current git state in a standard format (no Linear integration).

  npx skills add sorafujitani/skills/pr-generator

• review-code — Multi-agent code review. Four sub-agents review in parallel, each finding is checked for code existence and backed by official docs (WebSearch) to keep hallucinations out, then explained with prerequisite knowledge for the reader.

  npx skills add sorafujitani/skills/review-code

• exploratory-test — Detect the project's interface (CLI / API / Web API / GUI) and run an exploratory test pass end-to-end: plan → implement → execute → report.

  npx skills add sorafujitani/skills/exploratory-test

• property-based-test — Static analysis of the target picks property patterns, composes generators, generates the test code, runs it, and reports — across whatever PBT framework is in the project.

  npx skills add sorafujitani/skills/property-based-test

• karin-info — Pulls the latest news, releases, live dates and links about singer-songwriter Karin. via WebSearch / WebFetch.

  npx skills add sorafujitani/skills/karin-info

• local-repo-finder — Resolve another repository's path on the local machine, regardless of how the developer organises their workspace, using fd.

  npx skills add sorafujitani/skills/local-repo-finder

• doc-prerequisite-knowledge — Explain a topic from its prerequisite knowledge upward, with verified official-doc links at every step. Triggered by requests like 「〜について基本と前提知識から理解できるように教えて」 / "explain X from basics including prerequisites".

  npx skills add sorafujitani/skills/doc-prerequisite-knowledge

• graph-think-map — Turn prose-style or half-organized thoughts into a Mermaid graph plus a node/edge table and a five-axis validation report (logical consistency, concrete-vs-abstract balance, fact-vs-hypothesis separation, implicit premises, missing perspectives). Triggered by 「考えを整理して」「主張をマップして」 / "argument mapping" / "concept map".

  npx skills add sorafujitani/skills/graph-think-map

• skill-zip — Package a local Claude Code skill (~/.claude/skills/<name>/) into a ZIP at ~/Downloads/<name>.zip for upload to the Claude Desktop app's Settings → Customize → Skills. Bridges the gap between the CLI's directory-based skills and the desktop app's ZIP-upload flow. Triggered by 「skill を zip 化して」「Claude Desktop 用に package して」.

  npx skills add sorafujitani/skills/skill-zip

• ts-lint-searcher — Cross-search TypeScript lint rules across ESLint (with typescript-eslint), oxlint, and Biome lint. Given a behaviour to detect ("missing await on async calls") or a rule name (@typescript-eslint/no-floating-promises), returns a 3-way comparison table with rule names, support status, and official documentation links. A bundled mapping table answers well-known rules instantly; unknown rules trigger parallel WebFetch against the four official rule indexes. Triggered by 「<挙動> を検出する lint ルールは？」「<rule> の Biomeで同等のrule は？」 / "compare lint rules across ESLint / oxlint / Biome".

  npx skills add sorafujitani/skills/ts-lint-searcher

This repository lives directly at ~/.claude/skills/ on my own machine, which is the path Claude Code already loads user skills from. Editing a skill's SKILL.md here means the change is picked up by the next Claude Code session — no symlinks, no npx skills add round-trip.

1. Create <skill-name>/SKILL.md at the repo root with YAML frontmatter (name, description).
2. git add <skill-name>/ && git commit && git push.

Skill names follow the Anthropic Agent Skills naming conventions: lowercase letters, numbers, and hyphens only; gerund or short noun-phrase form; no vague labels (helper, utils, …) or reserved words (anthropic, claude).

Private skills stay in the same ~/.claude/skills/ directory but are listed in .gitignore and remain untracked. Add a directory name to .gitignore before its first git add to keep it local-only.

# Move any pre-existing private skills out of the way first if you have them.
git clone git@github.com:sorafujitani/skills.git ~/.claude/skills
# Drop your private skills back into ~/.claude/skills/. They remain untracked
# as long as their directory name is listed in .gitignore.

MIT — see LICENSE.