My Agent Skills.
| skill | Description |
|---|---|
| git-commit | Review the current changes and commit them in Conventional Commits format |
| git-fix-conflict | Detect and resolve conflicts from merge, rebase, cherry-pick, and more |
| git-squash | Squash and tidy the commits on the current branch, then push with force-with-lease when needed |
| github-fix-ci | Analyze CI failures and fix them automatically |
| github-issue-create | Verify the claims in the draft and search related issues, then create one GitHub Issue |
| github-issue-polish | Polish a GitHub issue until it can be implemented from the issue alone |
| github-issue-resolve | Run investigation → worktree creation → implementation → PR creation end-to-end from an issue |
| github-issue-update | Review open issues across the board and apply close / follow-up / label changes in bulk after approval |
| github-pr-create | Create a Pull Request from the current branch |
| github-pr-fix | Detect all PR problems (conflicts, CI failures, review comments) and fix them in a git worktree |
| github-pr-review | Review a PR with Finder/Verifier subagents and post a report and inline comments |
| github-resolve-pr-comment | Check PR review comments, address them, and reply |
| work-report | Report the PR, issue, review, and coding agent work done in a period as one HTML page |
github-pr-create is inspired by the pr skill in mattpocock/skills.
| skill | Description |
|---|---|
| claude | Call the Claude Code CLI non-interactively for consultation or delegated work |
| codex | Call the Codex CLI non-interactively for consultation or delegated work |
| resume-other-agent | Restore another coding agent's previous work from its session log and resume it |
| skill | Description |
|---|---|
| experiment-plan | Refine a machine learning experiment plan through dialogue and save it |
| grill-me | Keep questioning the user until every decision point in a plan or design is resolved |
| grill-self | Resolve every decision point through the agent's own research and self-questioning, then present a decision log |
| skill | Description |
|---|---|
| self-review | Adversarially review your own branch changes or one commit and report only verified findings without posting to GitHub |
| test-prune | Examine every test in the codebase and list the ones that can be deleted or merged |
The test-prune skill is based on catnose's post.
| skill | Description |
|---|---|
| skill-review | Score the quality of an agent skill (SKILL.md) and report findings with suggested fixes |
| writing-skills | Create or edit a skill test-first: watch subagents fail without it, then write it until they comply |
The writing-skills skill is adapted from the writing-skills skill in obra/superpowers v6.4.2. Its test-driven-development skill is bundled under references/ instead of being installed as a separate skill. The original is released under the MIT License, kept in skills/agent-skill/writing-skills/LICENSE.
| skill | Description |
|---|---|
| deep-research | Deep-research a topic with web search, present a cited Markdown note, and save it in the current directory on request |
| skill | Description |
|---|---|
| cognitive-rhythm-writing | Guidelines for designing rhythm and tempo in expository writing |
| japanese-tech-writing | Writing guidelines for Japanese technical documents and book manuscripts |
| japanese-writing | Write or revise Japanese prose under rules chosen for the kind of document |
| stop-ai-slop-jp | Rewrite AI-generated Japanese into natural, readable prose |
The cognitive-rhythm-writing skill is adapted from k16shikano's original gist.
The japanese-tech-writing skill is adapted from k16shikano's original gist. The original is released under the Unlicense.
The japanese-writing skill merges the rules of japanese-tech-writing and stop-ai-slop-jp, and draws on nanaism/yomiyasu.
| skill | Description |
|---|---|
| chat-note | Summarize research and discussion from the current conversation into a single Japanese Markdown file |
| doc-sync | Detect and fix drift between documentation and the implementation |
| exhtml | Explain a concept or system as a single self-contained HTML page with a glossary, figures, and a dark/light theme |
| exmd | Explain a concept or system as a single GitHub-flavored Markdown page with a glossary, table of contents, and Mermaid figures |
| skill | Description |
|---|---|
| duckdb | Check the schema of record-based data files with the DuckDB CLI before aggregating or converting them |
| wezterm-control | Control wezterm panes, tabs, and windows via wezterm cli |
flowchart LR
github-issue-resolve --> git-commit & github-pr-create
github-pr-create --> git-commit
github-pr-fix --> git-fix-conflict & github-fix-ci & github-resolve-pr-comment
cognitive-rhythm-writing --> japanese-tech-writing
exhtml -.-> japanese-writing & japanese-tech-writing & cognitive-rhythm-writing & stop-ai-slop-jp
exmd -.-> japanese-writing & japanese-tech-writing
Solid lines are required calls; dotted lines are optional calls, loaded when installed.
External tools each skill depends on are listed below. Without an optional tool, only the part that uses it is skipped. See each skill's README for details.
| Skill | Required | Optional |
|---|---|---|
git-commit |
git |
- |
git-fix-conflict |
git, gh |
- |
git-squash |
git, gh |
- |
github-fix-ci |
git, gh |
- |
github-issue-create |
gh |
- |
github-issue-polish |
git, gh |
- |
github-issue-resolve |
git, gh |
- |
github-issue-update |
gh |
- |
github-pr-create |
git, gh |
- |
github-pr-fix |
git, gh |
- |
github-pr-review |
git, gh, jq, python3 |
- |
github-resolve-pr-comment |
git, gh |
- |
work-report |
git, gh, jq, python3 (3.11+) |
- |
claude |
claude CLI |
- |
codex |
codex CLI |
- |
resume-other-agent |
git, rg, jq |
- |
experiment-plan |
- | oxfmt |
self-review |
git, gh, jq |
hunk, herdr |
skill-review |
shellcheck, shfmt, ruff |
claude CLI, codex CLI |
writing-skills |
- | Graphviz (dot) |
deep-research |
oxfmt |
- |
exhtml |
python3 |
- |
exmd |
python3 |
- |
duckdb |
duckdb CLI (1.5+) |
- |
wezterm-control |
wezterm CLI |
- |
The github-pr-review and self-review skills use the following three reviewer agents defined in agents/.
| Agent | Role |
|---|---|
code-reviewer-finder |
Looks across the whole change for candidate issues to resolve before it is finalized |
code-reviewer-contract |
Looks for gaps between the change and the spec's requirements, boundaries, and acceptance criteria. Launched only when a spec is resolved |
code-reviewer-verifier |
Tries to refute each candidate from finder and contract one at a time, keeping only the confirmed ones |
Agents ship through the Claude Code plugin and through apm. They do not specify per-agent tools restrictions; available tools and permissions follow each runtime's configuration. In OpenCode, mode: subagent makes them load as subagents. With apm install -t codex they are converted to Codex agent format. On other install methods the skills fall back to generic subagents given the same prompts.
- plugin: register mjun0812/skills as a plugin repository and install skills through it.
- npx skills / gh skills / apm: copy only the skills you want as plain files
See docs/skill-distribution-comparison.md for a detailed comparison of the five install methods.
# in session
/plugin marketplace add mjun0812/skills
/plugin install mjun-skills@mjun
# cli
claude plugin marketplace add mjun0812/skills
claude plugin install mjun-skills@mjun
# update
claude plugin update mjun-skills@mjuncodex plugin marketplace add mjun0812/skills
codex plugin add mjun-skills
codex plugin marketplace upgrade mjun
codex plugin remove mjun-skills# Select skills and target agents interactively
npx skills add mjun0812/skills
# Install all skills
npx skills add mjun0812/skills --all
# Install specific skills
npx skills add mjun0812/skills --skill git-commit --skill github-pr-create
npx skills add mjun0812/skills --skill git-commit -a claude-code -y
# Install to the user global scope instead of the project
npx skills add mjun0812/skills --skill git-commit -g
# Update (uses skills-lock.json)
npx skills updategh skill install mjun0812/skills
# Install all skills
gh skill install mjun0812/skills --all
# Install a specific skill with target agent and scope
gh skill install mjun0812/skills git-commit --agent claude-code --scope user
gh skill install mjun0812/skills git-commit --agent codex --scope project
# Pin to a git tag
gh skill install mjun0812/skills git-commit@v1.0.0
# Update
gh skill update# Install all skills
apm install mjun0812/skills
# Install specific skills (persisted to apm.yml / apm.lock.yaml)
apm install mjun0812/skills --skill git-commit
# Specify target harnesses
apm install mjun0812/skills -t claude,codex
# Pin to a git tag
apm install mjun0812/skills#v1.0.0
# Update
apm updateHow a skill removed from this repo is handled depends on the install method.
| Method | Behavior |
|---|---|
| Claude Code / Codex plugin | Removed automatically on update (the whole bundle is replaced) |
| apm | Removed automatically on update (cleaned up as stale files based on the lockfile) |
| npx skills / gh skill | Local copies remain; delete them manually |
Symlinks to every skill are committed in .claude/skills/ for Claude Code and .agents/skills/ for Codex. Start either tool in this repo to try the working-tree skills.
scripts/link_skills.py generates the symlinks. After a skill is added, removed, or renamed, the prek hook relinks the skills and stops the commit. Stage the changes and commit again. CI runs the same hook and catches missing links.
If the same skill is also installed in ~/.claude/skills/ or ~/.agents/skills/:
- Claude Code: the user copy takes precedence, so start with
claude --setting-sources project,local. This also skips the user CLAUDE.md, hooks, and permission settings. - Codex: both copies are listed, so move the skill out of
~/.agents/skills/while testing.
Releases are cut with the mise task:
mise run release 0.2.0It syncs the version into .claude-plugin/plugin.json and .codex-plugin/plugin.json, commits, tags v0.2.0, and pushes. The tag push triggers release.yml, which verifies that the tag matches the manifest versions and creates a GitHub Release with auto-generated notes.