Claude Prompts

What is Claude Prompts about?

Claude Prompts enables developers to manage, edit, and execute AI prompt templates directly from MCP‑compatible clients. Prompts can be hot‑reloaded, combined into multi‑step chains, wrapped with methodological frameworks, and guarded by quality gates, allowing autonomous or assisted AI workflows.

How to use Claude Prompts?

1. Installation – The simplest method is the NPX auto‑update configuration:

   {
     "mcpServers": {
       "claude-prompts": {
         "command": "npx",
         "args": ["-y", "claude-prompts@latest"]
       }
     }
   }
   

   Add the JSON to the MCP configuration file of your client (Claude Desktop, Cursor, Zed, etc.) and restart the client.
2. Quick start via Claude Code

   /plugin marketplace add minipuft/minipuft-plugins   # first‑time only
   /plugin install claude-prompts@minipuft
   >>tech_evaluation_chain library:'zod' context:'API validation'
   

3. Command syntax – Use the defined operators:
   • >>prompt_name – execute a prompt.
   • --> – pipe output to the next step (chain).
   • @Framework – inject a reasoning framework (e.g., @CAGEERF).
   • :: – attach a quality gate.
   • %modifier – toggle behavior such as %judge.
4. Custom resources – Override the default prompt directory with environment variables, e.g., MCP_RESOURCES_PATH=/my/prompts.
5. From source – Clone the repo, run npm install && npm run build, then point the MCP config to server/dist/index.js.

Key Features

• Hot Reload – Edit prompt files and see changes instantly without restarting the server.
• Chains – Compose multi‑step workflows using --> with automatic data passing.
• Frameworks – Built‑in structured thinking patterns (@CAGEERF, @ReACT, etc.).
• Gates – Inline quality criteria (:: 'cite sources') that Claude self‑validates.
• Judge Selection – %judge lets Claude pick the most appropriate tools for a task.
• Version History & Rollback – Track prompt revisions and revert to previous versions.
• Checkpoints – Git‑stash based snapshots for safe experimentation.
• Verification Gates (Ralph Loops) – Automated retry loops that run shell commands (e.g., npm test) until they pass.
• Cross‑client support – Works with Claude Desktop, Cursor, Gemini CLI, and any generic MCP client.

Use Cases

• AI‑assisted code review – Prompt chains that analyze, suggest fixes, and run tests automatically.
• Research summarisation – Chains that gather sources, enforce citation gates, and produce concise summaries.
• Documentation generation – Framework‑driven prompts that produce structured API docs with quality checks.
• Rapid prototyping – Hot‑reload prompts to iteratively refine assistant behavior.
• Enterprise policy drafting – Chains with verification loops to ensure compliance checks pass before final output.

FAQ

Q: Do I need Node.js installed? A: Yes, Node.js (v14+ recommended) is required for the NPX installation and for building from source.

Q: Can I use Claude Prompts with other LLM providers? A: The server is model‑agnostic; it works with any MCP‑compatible client, including Gemini via a separate extension.

Q: How do I add my own prompt files? A: Place them under a directory and set MCP_RESOURCES_PATH (or the more specific MCP_PROMPTS_PATH) in the server’s environment.

Q: What if a gate fails? A: The server can automatically retry, pause for user input, or abort based on the gate’s severity configuration.

Q: Is there a way to see prompt version history? A: Yes, use resource_manager(action:"history", id:"<prompt_id>") and resource_manager(action:"rollback", ...) to navigate versions.

Q: How do I contribute? A: Clone the repo, run npm install && npm run build && npm test. Submit pull requests following the CONTRIBUTING guidelines.

Claude Prompts MCP Server

Hot-reloadable prompts with chains, gates, and structured reasoning for AI assistants.

Quick Start • Features • Syntax • Docs

Quick Start

Claude Code (Recommended)

Step 1: Add the plugin marketplace (first time only)

/plugin marketplace add minipuft/minipuft-plugins

Step 2: Install the plugin

/plugin install claude-prompts@minipuft

Step 3: Try it

>>tech_evaluation_chain library:'zod' context:'API validation'

The plugin adds hooks that fix common issues:

Raw MCP works, but models sometimes miss the syntax. The hooks catch that. → hooks/README.md

User Data: Custom prompts stored in ~/.local/share/claude-prompts/ persist across updates.

Gemini CLI

The Gemini CLI extension is maintained in a separate repository for cleaner distribution:

# Install from the dedicated Gemini extension repo
gemini extensions install https://github.com/minipuft/gemini-prompts

The extension provides the same tools (prompt_engine, resource_manager, system_control) with Gemini-optimized hooks and context files.

→ Repository: github.com/minipuft/gemini-prompts

Works with the same prompts, gates, and methodologies as Claude Code. See the gemini-prompts README for hooks setup.

Claude Desktop

GitHub Release (recommended):

1. Download claude-prompts-{version}.mcpb from:
   → github.com/minipuft/claude-prompts/releases/latest
2. Drag into Claude Desktop Settings → MCP Servers
3. Done. Set custom prompts folder when prompted (optional).

The .mcpb bundle is self-contained (~5MB)—no npm install or node_modules required.

NPX (auto-updates):

// ~/Library/Application Support/Claude/claude_desktop_config.json (macOS)
// %APPDATA%\Claude\claude_desktop_config.json (Windows)
{
  "mcpServers": {
    "claude-prompts": {
      "command": "npx",
      "args": ["-y", "claude-prompts@latest"]
    }
  }
}

The npm package is also self-contained—all dependencies bundled, no postinstall required.

Restart Claude Desktop. Test it:

>>research_chain topic:'remote team policies' purpose:'handbook update'

Cursor

1-Click Install (from Cursor):

Note: The button above uses cursor:// protocol which works in Cursor's README preview but not on GitHub's web interface.

Manual Config (works everywhere):

1. Open Cursor Settings → MCP → Edit Config (or edit ~/.cursor/mcp.json)
2. Add this configuration:

{
  "mcpServers": {
    "claude-prompts": {
      "command": "npx",
      "args": ["-y", "claude-prompts@latest"]
    }
  }
}

3. Restart Cursor and test: resource_manager(resource_type:"prompt", action:"list")

Other MCP Clients

Add to your MCP configuration file:

{
  "mcpServers": {
    "claude-prompts": {
      "command": "npx",
      "args": ["-y", "claude-prompts@latest"]
    }
  }
}

For contributing or customizing the server itself:

git clone https://github.com/minipuft/claude-prompts.git
cd claude-prompts/server
npm install
npm run build   # Creates bundled dist/index.js (~4.5MB)
npm test        # Verify everything works

Point your MCP config to server/dist/index.js. The esbuild bundle is self-contained—no node_modules required at runtime.

Note: server/dist/ is not committed to git. You must run npm run build after cloning.

Transport options: --transport=stdio (default), --transport=streamable-http (for HTTP clients).

Custom Resources

Use your own prompts without cloning:

{
  "mcpServers": {
    "claude-prompts": {
      "command": "npx",
      "args": ["-y", "claude-prompts@latest"],
      "env": {
        "MCP_RESOURCES_PATH": "/path/to/your/resources"
      }
    }
  }
}

Your resources directory can contain: prompts/, gates/, methodologies/, styles/.

Priority: CLI flags > individual env vars > MCP_RESOURCES_PATH > package defaults.

See CLI Configuration for all options.

What You Get

🔥 Hot Reload

Edit prompts, test immediately. Better yet—ask Claude to fix them:

User: "The code_review prompt is too verbose"
Claude: resource_manager(action:"update", id:"code_review", ...)
User: "Test it"
Claude: prompt_engine(command:">>code_review")  # Uses updated version instantly

🔗 Chains

Break complex tasks into steps with -->:

analyze code --> identify issues --> propose fixes --> generate tests

Each step's output flows to the next. Add quality gates between steps.

🧠 Frameworks

Inject structured thinking patterns:

@CAGEERF Review this architecture    # Context → Analysis → Goals → Execution → Evaluation → Refinement
@ReACT Debug this error              # Reason → Act → Observe loops

🛡️ Gates

Quality criteria Claude self-checks:

Summarize this :: 'under 200 words' :: 'include key statistics'

Failed gates can retry automatically or pause for your decision.

✨ Judge Selection

Let Claude pick the right tools:

%judge Help me refactor this codebase

Claude analyzes available frameworks, gates, and styles, then applies the best combination.

📜 Version History

Every update is versioned. Compare, rollback, undo:

resource_manager(action:"history", id:"code_review")
resource_manager(action:"rollback", id:"code_review", version:2, confirm:true)

🔄 Checkpoints

Save working directory state before risky changes. Restore instantly if something breaks:

# Checkpoint before refactoring
resource_manager(resource_type:"checkpoint", action:"create", name:"pre-refactor")

# Something broke? Rollback to checkpoint
resource_manager(resource_type:"checkpoint", action:"rollback", name:"pre-refactor", confirm:true)

# List all checkpoints
resource_manager(resource_type:"checkpoint", action:"list")

Uses git stash under the hood. Pairs with verification gates for safe autonomous loops.

✅ Verification Gates (Ralph Loops)

Ground-truth validation via shell commands—Claude keeps trying until tests pass:

# You say this:
>>implement-feature :: verify:"npm test" loop:true

# Claude does this:
# 1. Implements feature
# 2. Runs npm test → FAIL
# 3. Reads error, fixes code
# 4. Runs npm test → FAIL
# 5. Tries again...
# 6. Runs npm test → PASS ✓
# You get working code.

Context Isolation: After 3 failed attempts, spawns a fresh Claude instance with session context—no context rot, fresh perspective, automatic handoff.

See Ralph Loops Guide for autonomous verification patterns and cost tracking.

Syntax Reference

Modifiers:

• %clean — No framework/gate injection
• %lean — Gates only, skip framework
• %guided — Force framework injection
• %judge — Claude selects best resources

Using Gates

# Inline (quick)
Research AI :: 'use recent sources' --> Summarize :: 'be concise'

# With framework
@CAGEERF Explain React hooks :: 'include examples'

# Programmatic
prompt_engine({
  command: ">>code_review",
  gates: [{ name: "Security", criteria: ["No hardcoded secrets"] }]
})

See Gates Guide for full schema.

Configuration

Customize via server/config.json:

The Three Tools

prompt_engine(command:"@CAGEERF >>analysis topic:'AI safety'")
resource_manager(resource_type:"prompt", action:"list")
resource_manager(resource_type:"checkpoint", action:"create", name:"backup")
system_control(action:"status")

How It Works

The feedback loop: Command with operators → Parse and inject methodology/gates → Claude executes and self-evaluates → Route: next step (PASS), retry (FAIL), or return result (done).

Documentation

• MCP Tooling Guide — Full command reference
• Prompt Authoring — Template syntax and schema
• Chains — Multi-step workflows
• Gates — Quality validation
• Ralph Loops — Autonomous verification patterns
• Architecture — System internals

Contributing

cd server
npm install
npm run build        # esbuild bundles to dist/index.js
npm test             # Run test suite
npm run validate:all # Full CI validation

The build produces a self-contained bundle (~4.5MB). server/dist/ is gitignored—CI builds fresh from source.

See CONTRIBUTING.md for workflow details.

License

MIT