PrePrompt + Cursor: Automatic Prompt Optimization via MCP

How PrePrompt Works in Cursor

When you submit a prompt in Cursor's Agent mode with PrePrompt installed, the following happens:

1. The UserPromptSubmit hook fires before the prompt reaches the LLM.
2. PrePrompt's heuristic classifier scores the prompt locally in under 1ms — no API call.
3. If the score is below 38, the prompt passes through untouched.
4. If the score is 38 or above, Claude Haiku rewrites the prompt to be more specific and actionable.
5. Cursor sends the (possibly rewritten) prompt to the LLM.
6. The interaction is logged locally to ~/.preprompt/history.db.

Stack memory compounds over time: as PrePrompt sees more of your prompts, it learns your preferred language, framework, and conventions, injecting them into future rewrites automatically.

Installation

1. Install PrePrompt

   pip install preprompt

2. Run the setup wizard

   This saves your Anthropic API key and registers PrePrompt with the IDEs it detects on your system.

   preprompt-install

   Alternatively, register Cursor specifically:

   python -c "from scripts.install_cursor import main; main()"

3. Verify the Cursor config was written

   cat ~/.cursor/mcp.json

   You should see a preprompt entry in the mcpServers object.

4. Restart Cursor

   MCP servers are loaded at startup. Restart Cursor to pick up the new configuration.

5. Test the integration

   In Cursor Agent mode, send a vague prompt like "write middleware that validates tokens and handles refresh". You should see a PrePrompt annotation in the terminal output, and Cursor will receive the rewritten prompt.

The MCP Configuration Entry

preprompt-install writes the following to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "preprompt": {
      "command": "python",
      "args": ["-m", "mcp_server.server"],
      "env": {
        "ANTHROPIC_API_KEY": "your-api-key-here"
      }
    }
  }
}

If you need to configure this manually, add the above entry to your existing ~/.cursor/mcp.json. If the file does not exist, create it with the full content above.

What Gets Optimized

PrePrompt's classifier routes each prompt through one of three paths:

• Pass — short prompts, factual questions ("what is JWT"), conversational responses ("thanks", "yes"). Zero latency, zero cost.
• Enrich — complex, vague, or multi-requirement prompts scoring 38+. Rewritten by Haiku. ~1–2 second latency, ~$0.001 cost.
• Clarify — very short vague prompts without technical context ("fix this", "improve it"). Returns a clarifying question to guide you toward a better prompt.

Frequently Asked Questions