agent-skills

A collection of Claude Code skills I use day-to-day. Drop any of them into a project's .claude/skills/ folder (or your global ~/.claude/skills/) to use them.

Skills

Skill	What it does
review-my-code	Senior-engineer review of uncommitted changes. Bugs, security, perf, missing tests, YAGNI/KISS. Designed to run automatically via a Stop hook.
add-bug	Log a bug to docs/bugs.md.
list-bugs	List bugs from docs/bugs.md with filters.
fix-bug	Pick a bug, fix it, test it, update its status.
add-feature	Log a feature request to docs/features.md.
playwright-cli	Automate the browser via playwright-cli for tests, screenshots, scraping.
excalidraw-diagram-skill	Generate Excalidraw diagrams that argue visually (not just boxes-and-arrows).

Install one skill (per-project)

git clone https://github.com/Crypto69/agent-skills.git /tmp/agent-skills
mkdir -p .claude/skills
cp -R /tmp/agent-skills/review-my-code .claude/skills/

Replace review-my-code with whichever skill you want. Restart Claude Code (or open a fresh session) and the skill is available — for example /review-my-code.

Install all skills (global)

git clone https://github.com/Crypto69/agent-skills.git /tmp/agent-skills
mkdir -p ~/.claude/skills
cd /tmp/agent-skills
for d in */; do cp -R "$d" ~/.claude/skills/; done

Auto-run review-my-code after every Claude turn

The flagship use case — every time Claude finishes a turn, a fresh subagent reviews the uncommitted diff before you commit it.

Add this to your project's .claude/settings.json (or ~/.claude/settings.json for global):

{
  "hooks": {
    "SessionStart": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "sid=$(jq -r .session_id); { git rev-parse HEAD; git diff HEAD; git status --porcelain; } 2>/dev/null | shasum > \"/tmp/cc-review-$sid\""
          }
        ]
      }
    ],
    "Stop": [
      {
        "hooks": [
          {
            "type": "command",
            "command": "sid=$(jq -r .session_id); cur=$({ git rev-parse HEAD; git diff HEAD; git status --porcelain; } 2>/dev/null | shasum); saved=$(cat \"/tmp/cc-review-$sid\" 2>/dev/null); [ \"$cur\" != \"$saved\" ] || exit 1"
          },
          {
            "type": "agent",
            "prompt": "Invoke the review-my-code skill on the git diff of changes made in this session. Report findings concisely.",
            "model": "claude-sonnet-4-6",
            "timeout": 180,
            "statusMessage": "Reviewing changes..."
          }
        ]
      }
    ]
  }
}

How it works:

• SessionStart snapshots the working-tree state to /tmp/cc-review-<session_id> (HEAD commit + git diff HEAD + untracked files, hashed).
• Stop precheck re-hashes the tree at end-of-turn. If it matches the snapshot, exit 1 → agent is skipped. Plan-only sessions, exploration, research, and Q&A all cost zero tokens.
• Stop agent runs only when the tree actually changed during this session — pre-existing uncommitted clutter doesn't trigger reviews.

Field	Purpose
SessionStart command	Snapshot working-tree hash, keyed by session id.
Stop precheck	Compare current tree to snapshot; exit 1 to skip review.
type: "agent"	Spawn a fresh subagent (separate context) to run the review.
prompt	What the subagent is asked to do.
model	Reviewer model. Sonnet is fast + cheap; switch to Opus for deeper reviews.
timeout	Seconds before the hook is killed (180s handles most diffs).
statusMessage	Shown in the Claude Code status line while reviewing.

Requires jq on PATH (preinstalled on macOS, available via apt/brew elsewhere).

After editing settings, open /hooks once or restart Claude Code so the watcher picks up the new config.

See review-my-code/README.md for full details.

Uninstall

rm -rf .claude/skills/<skill-name>            # or ~/.claude/skills/<skill-name>

If you added the Stop hook, remove the Stop block from your settings.json too.

License

MIT — see LICENSE.