Each document narrows the one above it until a subtask is small enough to be a single commit, and execution stops for review before committing - so the plan stays legible to you instead of disappearing into one long generation.
An MCP server that gives an agent a fixed development workflow, with a Rails leaning. It collects functional requirements, turns them into a PRD, splits each feature into a pull-request-sized FRD, then breaks that into a task design document where every subtask is one commit. Documents and task state are written into .cursor/scratch/ in the repository.
- Project requirements collected through a structured interview - purpose, users, capabilities, workflows, edge cases
- A PRD generated from those requirements, broken into features and user stories
- A feature requirements document sized to a single pull request
- An implementation plan whose subtasks map one-to-one onto commits
- Architecture documentation for the project
- Those subtasks set up as trackable work
- One task executed at a time, pausing for review before the commit
An MCP client with an mcp.json, and a repository the server can write to - the documents, the task list and the progress notes all land in .cursor/scratch/. For a system you already understand, skip the first three documents and start at the task design step.
One command — docker run -i --rm cestbalez/railagent:latest