AI Tools

Lightweight, language-agnostic skills, hooks and statusline for Claude Code / Copilot CLI / Gemini CLI. Designed for teams working across multiple projects and tech stacks.

Skills

Skill	Command	Description
work-through-this	/work-through-this [task]	Entry point — analyzes task type, routes to the right skill, asks if ambiguous
brainstorm	/brainstorm [idea]	Reads project context, asks questions, proposes approaches, writes spec/plan
implement	/implement [path/to/plan.md]	Execute plan via subagents with verification after each step
debug	/debug [bug description]	Reproduce first, then hypothesize → test → fix → verify
review	/review [PR number / branch]	Full-diff code review with file:line references, severity levels, test coverage checks

Installation

git clone <repo-url>
cd ai-tools
./install.sh

Creates symlinks in ~/.claude/skills/ (Claude Code) and ~/.agents/skills/ (Copilot CLI, Codex, Gemini CLI).

Optional extras — run from inside the cloned repo via claude:

What	Command	What it does
Notification hooks	/install-hooks	Adds Stop + Notification hooks to ~/.claude/settings.json
Statusline	/install-statusline	Symlinks cli/statusline.sh to ~/.claude/ and adds statusLine to settings

Both extras back up your settings.json before touching it, preserve all other fields, and have matching /uninstall-* commands.

Update

git pull && ./install.sh

Skills are symlinks, so git pull is enough — no re-install needed. Hooks and statusline re-install only if hooks.json, statusline.json, or cli/statusline.sh change.

Usage

Start with the entry point:

/work-through-this add user authentication

Or call any skill directly:

/brainstorm new caching layer for API responses
/implement docs/plans/2026-04-01-caching-plan.md
/debug login fails after token refresh
/review 42

How It Works

• work-through-this analyzes your task and routes to the appropriate skill. If task is ambiguous (bug + feature) — asks which to tackle first
• brainstorm reads project's CLAUDE.md first, adapts to task size: small tasks get a plan directly, large tasks get a spec first
• implement dispatches one subagent per task step with verification after each, then a review-subagent checks everything at the end
• debug reproduces the bug first, then tests hypotheses one at a time, most probable first
• review reads the full diff and reports issues by severity (critical / important / nit), flags missing test coverage

Hooks

Notification hooks for Claude Code (macOS). Requires terminal-notifier and jq:

brew install terminal-notifier jq

Inside the cloned repo:

claude
/install-hooks      # merge Stop + Notification hooks into settings.json
/uninstall-hooks    # remove them

What you get:

• Stop — plays Glass sound and shows a notification when Claude finishes responding
• Notification — plays Ping sound and shows the message when Claude is waiting for input

Manual install — copy the hooks object from hooks.json into your ~/.claude/settings.json by hand.

Statusline

Custom statusline with model, git branch and context-window progress bar. Requires jq:

brew install jq

Inside the cloned repo:

claude
/install-statusline      # symlinks cli/statusline.sh and adds statusLine to settings
/uninstall-statusline    # removes the symlink and settings entry

The install symlinks cli/statusline.sh into ~/.claude/statusline.sh, so editing the script in the repo updates everywhere. Restart Claude Code to see changes.

Project Verification

Skills that run verification (implement, debug) look for a Verification section in your project's CLAUDE.md. Example:

## Verification
- `npm test`
- `npm run lint`

If no Verification section is found, skills auto-detect the language and run standard checks:

Project marker	Language	Commands
pubspec.yaml	Dart/Flutter	dart analyze, flutter test
package.json	JS/TS	npm test, npx tsc --noEmit
build.gradle	Kotlin/Java	./gradlew build
pyproject.toml / setup.py	Python	pytest
go.mod	Go	go build ./..., go test ./...