Claude Code 配方集

Claude Code Recipes

49 copy-paste recipes for Claude Code - commands, subagents, hooks, skills, and MCP configs that work out of the box. Install them as plugins in two commands, or copy the files you want.

Stop configuring from scratch. Drop these recipes into your .claude/ folder and start building.

Something installed but silent? Claude Code config fails quietly, so start at TROUBLESHOOTING.md.

Install as Plugins

The repo is also a Claude Code plugin marketplace. Inside Claude Code:

/plugin marketplace add Sagargupta16/claude-code-recipes
/plugin install recipe-commands@claude-code-recipes

On Claude Code 2.1.275 or later, one line does both:

/plugin install recipe-commands --marketplace Sagargupta16/claude-code-recipes

Each plugin is one recipe type, so you take only what fits:

scripts/build_marketplace.py generates .claude-plugin/marketplace.json from the files on disk, and CI fails if the two drift, so a new recipe cannot miss the plugins. CI also runs claude plugin validate --strict on it. MCP configs and CLAUDE.md templates stay copy-paste: they need your own connection strings and project details.

How to Use

1. Browse the recipe catalog below
2. Copy the file you need
3. Paste it into your project's .claude/ directory:
   • Commands go in .claude/commands/
   • Subagents go in .claude/agents/
   • Skills go in .claude/skills/
   • Hooks go in .claude/settings.json
   • MCP configs go in .mcp.json
4. Use it immediately - every recipe works out of the box

Where Each Recipe Plugs In

Every recipe type hooks into a different point of a Claude Code session. Knowing which one fires when is most of the debugging when a recipe stays silent.

flowchart TD
    subgraph start["Session start"]
        CM[CLAUDE.md<br/>templates] --> CTX[Session context]
        MCP[.mcp.json<br/>MCP configs] --> TOOLS[Extra tools]
        SK[.claude/skills<br/>skill descriptions] --> CTX
    end
    U[You type a prompt<br/>or /command] --> CMD[.claude/commands<br/>prompt + model]
    CMD --> LOOP((Agent loop))
    CTX --> LOOP
    TOOLS --> LOOP
    LOOP -- delegates --> AG[.claude/agents<br/>subagent with scoped tools]
    LOOP -- loads on match --> SK
    LOOP -- tool call --> PRE{PreToolUse hook}
    PRE -- exit 2 blocks --> LOOP
    PRE -- allowed --> RUN[Bash / Edit / Write]
    RUN --> POST[PostToolUse hook<br/>format the edited file]
    POST --> LOOP
    LOOP -- waiting on you --> NOTE[Notification hook]

Recipe Catalog

Commands (15 recipes)

Drop these into .claude/commands/ and use them with /command-name.

Subagents (9 recipes)

Drop these into .claude/agents/ for specialized delegation.

Hooks (7 recipes)

Event-triggered scripts for automation.

Skills (5 recipes)

Reusable knowledge modules - drop the folder into .claude/skills/.

MCP Configs (5 recipes)

Model Context Protocol server configurations for .mcp.json.

Workflows (5 recipes)

End-to-end guides combining multiple recipes.

CLAUDE.md Templates (3 recipes)

Optimized project configuration templates.

Quick Reference

See the cheatsheet for a one-page summary, and TROUBLESHOOTING.md when a recipe is installed but does nothing.

Verifying a Recipe

Every recipe in this repo is checked on each push and pull request: frontmatter validity, the hook config schema, JSON parsing, shellcheck, working links, that the recipe counts above match the files on disk, and that the plugin marketplace matches the recipes.

The hooks are also tested for what they do, not only for how they are written. scripts/test-hooks.sh replays hook payloads against every script and asserts the exit code and stderr Claude would see: 77 checks, run on Ubuntu, macOS and Windows. Linters, test runners and notifiers are stubbed on PATH, so a case can force a lint failure or a hostile notification message without installing anything. Restore the 1.1.0 notification.sh or pre-push-test.sh and the suite fails on the injection case and the pytest-exit-5 case.

Run the same checks yourself:

bash scripts/validate-recipes.sh
bash scripts/test-hooks.sh

Installation

One recipe at a time

# Copy a command
cp claude-code-recipes/commands/code-review.md .claude/commands/

# Copy a subagent
cp claude-code-recipes/subagents/researcher.md .claude/agents/

# Copy a skill
cp -r claude-code-recipes/skills/react-patterns .claude/skills/

All recipes at once

# Clone and copy everything
git clone https://github.com/Sagargupta16/claude-code-recipes.git /tmp/recipes
mkdir -p .claude/commands .claude/agents .claude/skills
cp /tmp/recipes/commands/*.md .claude/commands/
rm -f .claude/commands/README.md
cp /tmp/recipes/subagents/*.md .claude/agents/
rm -f .claude/agents/README.md
cp -r /tmp/recipes/skills/* .claude/skills/
rm -f .claude/skills/README.md

Contributing

We welcome recipe contributions! See CONTRIBUTING.md.

Ways to contribute:

• Add a recipe: Commands, subagents, skills, hooks, or MCP configs
• Improve existing recipes: Better prompts, additional edge cases
• Add a workflow: End-to-end guides for common scenarios
• Report issues: Recipes that don't work as expected

Related

License

MIT - use these recipes however you want.