Agentic Layer Starter Kit

A hands-on starter kit for engineers who want to unlock the full power of Claude Code.

Stop using Claude Code as a fancy autocomplete. Start building your own agentic layer — custom commands, skills, hooks, and agents that compound your productivity over time.

This repo gives you working examples to learn from, a structure to build on, and a launchpad for making Claude Code truly yours.

---

Why This Exists

Claude Code ships with powerful extension points — commands, skills, hooks, and agents — but most engineers never touch them. The docs explain what they are, but not how to think about them or where to start.

This starter kit bridges that gap. It includes:

• Real, working examples you can run today
• A clean directory structure that follows Claude Code conventions
• Room to grow so you can add your own tools as you learn

The goal is simple: get comfortable, experiment, and iterate your way toward an agentic layer that fits how you work.

---

What's Inside

.claude/
├── commands/          # Slash commands you can invoke with /<name>
│   ├── prime.md           → /prime — Orient yourself in any codebase
│   ├── commit-pro.md      → /commit-pro — Professional conventional commits
│   └── youtube-transcribe.md → /youtube-transcribe — Extract video transcripts
│
├── skills/            # Complex, multi-step capabilities
│   └── deep-research/
│       └── SKILL.md       → Deep parallel research across multiple sources
│
├── hooks/             # Event-driven automations
│   └── git_safety_guard.py  → Blocks destructive git/filesystem commands
│
├── settings.json      # Hook wiring and project-level config
│
└── agents/            # Custom agent definitions (your turn)

ai_docs/               # Reference material loaded as context
├── README.md              → URLs to fetch as runtime context
└── backend-project-ideas.md → 20 portfolio-ready project ideas

---

Included Examples

Commands

Command	What It Does
/prime	Runs git ls-files, reads the README, and produces a codebase summary. A fast way to orient yourself in any project.
/commit-pro	Analyzes your staged and unstaged changes, generates a conventional commit message, and offers workflow options — new branch, push to main, or create a PR.
/youtube-transcribe	Downloads a YouTube video's transcript (or generates one via Whisper), cleans it up, and produces a markdown summary with key takeaways.

Skills

Skill	What It Does
deep-research	Launches 4–6 parallel research agents that investigate a topic from different angles — architecture, best practices, comparisons, case studies. Synthesizes everything into a structured markdown report saved to research/.

Hooks

Hook	What It Does
git_safety_guard.py	A PreToolUse hook that intercepts every Bash command before execution. Blocks destructive operations — git reset --hard, git push --force, rm -rf, git clean -f, and more — while allowing safe variants like git checkout -b and rm -rf /tmp/. Wired up in settings.json.

Reference Docs

The ai_docs/ directory holds markdown files that provide Claude with additional context. The included backend-project-ideas.md contains 20 intermediate-level REST API project ideas with database schemas and a decision matrix — useful when you need inspiration for your next build.

---

Getting Started

Prerequisites

• Claude Code CLI installed and authenticated

Usage

1. Clone the repo

   git clone https://github.com/your-username/Agentic_Layer_Starter_Kit.git
   cd Agentic_Layer_Starter_Kit

2. Launch Claude Code

   claude

3. Try a command

   /prime
   

   Claude will read the codebase and give you a structured summary.

4. Explore the examples — read through the files in .claude/commands/, .claude/skills/, and .claude/hooks/ to understand how they work.

5. Start building your own — add a new .md file to commands/, define a skill in skills/, or wire up a hook.

---

Build Your Own

Adding a Command

Create a markdown file in .claude/commands/:

.claude/commands/my-command.md

The file contents become the prompt. Use sections, instructions, and tool references to define what the command does. Invoke it with /my-command.

Adding a Skill

Create a directory in .claude/skills/ with a SKILL.md file:

.claude/skills/my-skill/SKILL.md

Skills are more complex than commands — they can define multi-step workflows, spawn parallel agents, and produce structured output.

Adding a Hook

Hooks are scripts that run in response to Claude Code events (like PreToolUse, PostToolUse, etc.). Place your script in .claude/hooks/ and wire it up in .claude/settings.json. See git_safety_guard.py for a working example — it reads the tool input from stdin as JSON and exits with a deny decision to block dangerous commands.

Adding an Agent

Custom agents go in .claude/agents/. These define specialized sub-agents with specific tool access and behaviors.

---

Directory Conventions

Directory	Purpose
.claude/commands/	Slash commands — simple, user-invoked prompts
.claude/skills/	Skills — complex, multi-step capabilities
.claude/hooks/	Hooks — event-driven automations
.claude/agents/	Agents — specialized sub-agents
ai_docs/	Reference docs loaded as context
research/	Output from the deep-research skill

---

Philosophy

Your agentic layer should evolve with you. Start small:

1. Use the included examples to understand the patterns
2. Modify them to fit your workflow
3. Create new ones as you discover repetitive tasks
4. Share what works with your team

The best agentic layer is the one you actually use. Don't over-engineer it — grow it organically from real needs.

---

Contributing

Found a useful command or skill? Contributions are welcome. Open a PR with your addition and a brief description of what it does and why it's useful.

---

License

MIT License — see LICENSE for details.