TL;DR

Currently, the prompts used by the Gemini CLI are defined inline within the workflow files.

When we improve a prompt, users who have copied our provided workflows won't get the updates unless they manually copy the new prompt from our repository. This means users can easily be running on old, sub-optimal prompts.

Detailed design

Consider serving prompts via the Model Context Protocol (MCP), similar to how we currently serve custom tools.

This would involve:

1. Defining our prompts in a structured format that the MCP server can understand.
2. Configuring the MCP server to expose these prompts to the Gemini CLI.
3. Updating the workflows to reference the prompts via MCP.

Additional information

Learn more in https://modelcontextprotocol.io/specification/2025-06-18/server/prompts