A Zero-Dependency Stateful PRD Framework for AI-Driven Development

npx bivvy init --cursor

Then ask your AI agent to create a new climb and you're ready to go!

**(NOTE: We suggest you commit the created Bivvy files before making additional changes)

Currently, Bivvy supports:

• Cursor (✅ Available now)
• Windsurf (🚧 Coming soon)

Want to see Bivvy support another client? Open an issue!

Bivvy provides a structured framework for AI-driven development through a combination of Product Requirements Documents (PRDs) and task management. Here's how it works:

When you run bivvy init --cursor, Bivvy:

1. Creates a .cursor/rules/bivvy.mdc file with the AI interaction rules
2. Sets up a .bivvy directory with example files
3. Creates a .bivvy/complete directory for finished work

A "Climb" is Bivvy's term for a development project, which can be a feature, bug fix, task, or exploration. Each Climb consists of two key components:

1. PRD (.bivvy/[id]-climb.md)

   • Contains the project requirements and specifications
   • Includes metadata like ID, type, and description
   • Documents dependencies, prerequisites, and relevant files
   • Structured as a markdown file with YAML frontmatter

2. Moves (.bivvy/[id]-moves.json)

   • A JSON file containing the task list
   • Each move has a status: todo, climbing, skip, or complete
   • Moves can be marked with rest: true for mandatory checkpoints
   • Tasks are executed in strict order

.bivvy/
├── [id]-climb.md      # Active PRD
├── [id]-moves.json    # Active task list
└── complete/          # Completed climbs
    ├── [id]-climb.md
    └── [id]-moves.json

1. Starting a Climb

   • Ask your AI agent to create a new climb
   • The agent will help you draft the PRD
   • You must approve the initial PRD before proceeding

2. During Development

   • The agent works through moves one at a time
   • Stops for approval at rest: true checkpoints
   • Skips tasks marked as skip (can return to them later)
   • Updates move statuses in real-time

3. Completing a Climb

   • When all moves are done, the agent will:
     • Ask to mark the climb as complete
     • Move files to .bivvy/complete/
     • Stop tracking the climb

1. Initialize Bivvy: npx bivvy init --cursor
2. Start a new climb: "Create a new feature climb for adding user authentication"
3. Review and approve the PRD
4. Review and approve the moves list
5. Let the agent work through the moves
6. Review and approve each significant change
7. Complete the climb when done

• Always review PRDs and moves lists before approval
• Use rest: true for important checkpoints
• Mark non-critical tasks as skip to focus on core functionality
• Keep moves small and focused (2-3 code changes)
• Trust the process and follow moves in order

Want to try out changes or contribute? Clone the repo and copy the files manually:

git clone https://github.com/taggartbg/bivvy.git
mkdir your-project/.bivvy
cp -r bivvy/src/example/* your-project/.bivvy
cp bivvy/src/rules/.cursor.mdc your-project/.cursor/rules/bivvy.mdc

While Bivvy is a small project, I'd love to see it grow! Contributions are welcome through:

• Pull Requests for new features or improvements
• Issues for bug reports or feature requests
• Discussions in the Issues section

Let's make AI-driven development more structured and efficient together!

MIT