Prompt Clarifier — MCP Agent

Stop the back-and-forth with your LLM. This agent asks the right questions before executing your request.

How it works

You write: "Create an ONNX model without opset"
                    ↓
The MCP server detects the domain and returns a system prompt
                    ↓
Your IDE's own LLM asks targeted questions (no external LLM)
                    ↓
You answer (or type "go" to stop)
                    ↓
An enriched, precise prompt is generated
                    ↓
Your LLM produces exactly what you wanted — on the first try

Architecture

The MCP server is LLM-agnostic: it makes no external API calls and requires no API key.

| Responsibility | Who handles it | | --- | --- | | Domain detection | MCP server | | Session state (Q&A history) | MCP server | | Question generation | Your IDE's LLM (Junie, Cursor, Claude Code, Copilot…) | | Knowledge base search | Your IDE's LLM (Confluence, Jira, Figma, Notion, Linear, GitHub Issues…) |

Call flow:

• First call (prompt) — the server creates a session, detects the domain, and returns a system_prompt + user_message that your IDE injects into its LLM to generate the first question.
• Subsequent calls (session_id + answer) — the server saves the answer and returns an updated system_prompt. After 5 answers or the word "go", it returns the enriched final_prompt.

Automatically detected domains

| Domain | Detected patterns | Question focus | | --- | --- | --- | | ml_onnx | onnx | runtime, opset, tensor shapes, operators | | ml_h2o | h2o, automl, gbm, mojo, pojo | algorithm type, target variable, export format | | data_table | datatable, schema, dataframe, column | schema, data types, volume, use case | | ml_general | pytorch, tensorflow, sklearn, model… | framework, task type, data format | | general | (everything else) | objective, environment, existing context |

Knowledge base search

The returned system prompt instructs your IDE's LLM to use its connected MCP tools to gather internal context before formulating questions. This produces questions tailored to your organization's actual standards and workflows rather than generic ones.

Supported platforms (if connected in your IDE):

| Platform | MCP tool used | What is searched | | --- | --- | --- | | Confluence (Atlassian Rovo) | searchConfluenceUsingCql | Internal standards, architecture decisions, naming conventions | | Jira (Atlassian Rovo) | searchJiraIssuesUsingJql | Open issues, epics, current constraints | | Figma | Figma MCP tool | Design specs, component names | | Notion | Notion MCP tool | Internal documentation | | Linear | Linear MCP tool | Open issues, roadmap | | GitHub Issues | GitHub MCP tool | Open issues, discussions |

If none of these tools are connected, the LLM falls back to domain best practices.

Installation

No API key required. The server makes no external LLM calls.

Option 1 — Global install (recommended)

npm install -g prompt-clarifier-mcp

Then use prompt-clarifier-mcp as the command in your IDE config instead of npx:

{
  "command": "prompt-clarifier-mcp",
  "args": []
}

Option 2 — On-demand via npx

No install needed. Use npx -y prompt-clarifier-mcp directly in your IDE config (see examples below). The package is fetched automatically on first run.

Cursor

Open ~/.cursor/mcp.json (create it if it does not exist):

{
  "mcpServers": {
    "prompt-clarifier": {
      "command": "npx",
      "args": ["-y", "prompt-clarifier-mcp"]
    }
  }
}

Restart Cursor. The clarify tool is now available.

Claude Desktop

Open the config file:

• Windows: %APPDATA%\Claude\claude_desktop_config.json
• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "prompt-clarifier": {
      "command": "npx",
      "args": ["-y", "prompt-clarifier-mcp"]
    }
  }
}

Restart Claude Desktop.

VS Code (with GitHub Copilot or Continue)

Create or open .vscode/mcp.json at the root of your project:

{
  "servers": {
    "prompt-clarifier": {
      "command": "npx",
      "args": ["-y", "prompt-clarifier-mcp"]
    }
  }
}

IntelliJ IDEA / PyCharm (2025.1+)

1. Open Settings → Tools → AI Assistant → Model Context Protocol (MCP)
2. Click + to add a new server
3. Choose "Command" as the type
4. Fill in:
   • Name: prompt-clarifier
   • Command: npx
   • Arguments: -y prompt-clarifier-mcp
5. Click OK and restart the IDE

Enable "Codebase" mode in the AI Assistant chat for MCP tools to be available.

Claude Code

claude mcp add prompt-clarifier npx -- -y prompt-clarifier-mcp

Usage

In any IDE chat, call the clarify tool with your prompt:

Use the clarify tool with this prompt: "Create an ONNX model without opset"

The server will:

1. Detect the domain (ml_onnx in this example)
2. Return a system prompt your LLM uses to ask targeted questions
3. Record each answer in the session
4. Generate the enriched prompt after 5 answers or as soon as you type "go"

To stop questions at any time, simply write:

• go / commence / start / proceed / just do it / enough

Response format

First call:

{
  "session_id": "uuid",
  "system_prompt": "...",
  "user_message": "Here is the user prompt to clarify: ...",
  "instructions": "Ask the first clarifying question now. Include the session_id..."
}

Subsequent calls:

{
  "session_id": "uuid",
  "system_prompt": "...",
  "user_message": "The user answered: ... Ask the next question.",
  "qa_count": 2
}

Pass back the question field (the question your LLM just asked) alongside answer so the session history stays complete and the enriched prompt includes the full Q&A context.

Final response:

{
  "final_prompt": "Initial prompt\n\n## Additional context collected\n..."
}

Requirements

• Node.js 18 or higher
• An IDE with MCP support (see list above)

Local development

git clone https://github.com/Didou555/prompt-clarifer
cd prompt-clarifer
npm install
npm run build

# Test with the MCP inspector
npx @modelcontextprotocol/inspector node dist/index.js

Contributing

Contributions are welcome! In particular:

• New domain detection patterns (src/clarifier.ts → detectDomain)
• New stop keywords in other languages (GO_KEYWORDS)
• New question angles per domain (DOMAIN_ANGLES)

License

MIT