Claude Skills Repository

Personal Claude Code skills for OlliTapio.

Leak guard (personal repo — no work content)

.githooks/ blocks commits and pushes that contain work-related terms or use a non-personal email (allowed: olli@otl.fi, the GitHub noreply, Claude's co-author address). Terms live in the gitignored .githooks/denylist.local (one case-insensitive regex per line) so they never get committed; the hooks fail closed if it's missing. After cloning:

git config core.hooksPath .githooks
git config user.email olli@otl.fi
# then create .githooks/denylist.local

Never bypass with --no-verify. Generalize real-world examples so no employer, client, product, or internal repo name appears — see .claude/rules/generic-content.md (always on).

Claude Code hooks (.claude/settings.json) add a layer on top, since the denylist only knows listed terms:

• Every file edit — leak-guard.sh text checks it against the denylist, and a prompt hook has a model reject real employer/client/product/person names. A block is fed back to Claude to generalize.
• git push — .claude/hooks/review-push.sh runs the denylist plus a claude -p review over all outgoing commits (messages + diffs) and blocks the push on any hit. Fails closed.

Writing Skills

See .claude/rules/skill-authoring.md (auto-loads when editing skills/**/SKILL.md).

Updating Skills

When modifying skills in this repo:

1. Edit the skill files directly in skills/<skill-name>/SKILL.md
2. After making changes, commit and push to this repo:

   git add -A && git commit -m "Update <skill-name>: <brief description>" && git push

The skills are symlinked from ~/.claude/skills/ to this repo, so changes take effect immediately.

Adding New Skills

1. Create a new directory: skills/<skill-name>/
2. Add SKILL.md with the skill definition
3. Create the symlink: ln -s /Users/olli/repositories/claude_skills/skills/<skill-name> ~/.claude/skills/<skill-name>
4. Commit and push

Installing as a plugin (cloud sessions)

This repo is also a plugin marketplace (.claude-plugin/), so cloud sessions can load the skills. Enable it in a project's .claude/settings.json:

{
  "extraKnownMarketplaces": {
    "olli-skills": { "source": { "source": "github", "repo": "OlliTapio/claude_skills" } }
  },
  "enabledPlugins": { "olli-skills@olli-skills": true }
}

Plugin skills are namespaced (/olli-skills:pr). plugin.json has no version, so installs track the latest commit. Skills that read repo files (templates/, references/) must use paths relative to the skill's base directory, never ~/.claude/skills/... or a local checkout path. Validate with claude plugin validate .. Locally, keep using the symlinks rather than also enabling the plugin, or every skill shows up twice.

Current Skills

• plan - TDD-based planning workflow
• pr - Create PRs with quality checks, review guideline validation, and relevance check
• pr-review - Review GitHub PRs with severity-ranked findings; auto-fixes P0/P1 and has the reviewing agents confirm
• explain-diff - Rich explanation of a code change/diff/PR as an HTML file; adapted from Geoffrey Litt's gist
• finalize - Create PR, review, fix issues, squash merge end-to-end
• test-on-main - Exit worktree, checkout branch on main, merge main, run tests
• frontend-design - Distinctive, production-grade frontend interfaces
• analytics-view - Build analytics surfaces with single-source-of-truth metrics and cross-surface parity tests
• rule-authoring - Write path-scoped or always-on .claude/rules/ files following Anthropic best practices
• quality-gates - Set up a project with language-agnostic code-smell/quality gates via git pre-commit (fast) and pre-push (slow) hooks
• setup-project - Scaffold a new project: LLM-friendly language, layered architecture (controllers/services/repos/views), .claude rules, quality gates

Global hooks (global-hooks/)

Hooks registered in the user-global ~/.claude/settings.json, so they apply to every repo on the machine. Symlinked from this repo (like skills), not copied — unlike templates/hooks/, which are project-level and copied into one repo.

• fetch-default-branch.sh - fetches origin/<default-branch> before Claude Code cuts a worktree, so worktree.baseRef: fresh isn't resolved against a stale local ref; also fast-forwards the local default branch (what agents read via git log main / git merge main) when that is provably lossless

Installing one takes two steps — symlink and a settings.json registration:

mkdir -p ~/.claude/hooks
ln -s /Users/olli/repositories/claude_skills/global-hooks/<hook>.sh ~/.claude/hooks/<hook>.sh
# then merge the hook's event registration into ~/.claude/settings.json

See global-hooks/README.md for the events, the JSON to merge, and design constraints.

Templates

Reusable configuration templates in templates/:

Hooks (templates/hooks/)

Pre-configured Claude Code hooks for quality enforcement:

# Copy to your project
mkdir -p .claude/hooks
cp templates/hooks/settings.json .claude/
cp templates/hooks/*.sh .claude/hooks/
chmod +x .claude/hooks/*.sh

• precommit.sh - Blocks git commit if checks fail
• lint-fix.sh - Auto-fixes lint after file edits

See templates/hooks/README.md for configuration.

Git hooks (templates/git-hooks/)

Versioned git hooks (run outside Claude too), installed by the quality-gates skill:

• pre-commit - fast, staged files: format, lint, complexity, duplication
• pre-push - slow, whole repo: types, tests, dead-code

See templates/git-hooks/README.md.

Rules (templates/rules/)

Reusable .claude/rules/*.md files installed by the setup-project skill:

• layered-architecture.md - controllers/services/repos/views responsibilities and dependency direction