These preferences apply globally to all projects.
- Always work on branches, not main - Create a feature branch before making changes
- Create PRs at reviewable milestones - Not one mega-PR at the end
- Use semantic commits -
feat:,fix:,refactor:, etc.
- Use beads for tracking, not markdown plans - Plans as committed markdown create PR noise and get stale
- Epic bead for features - Store design decisions in
--designfield - Child beads for meaningful chunks - 3-7 per feature, each could be a commit or handoff point
- Update
--notesas work progresses - Capture implementation decisions - Use TodoWrite for atomic steps - Within a single work session only
- Explain what you're doing and why, then act - Don't ask permission for each step
- Offer a ripcord - "Say 'stop' if you want to change approach"
- Be smart about parallelization - Spin up subagents for independent work, sequential for dependencies
- Sense when context is running low - Proactively suggest handoff before it's a problem
- First-pass review is optional - Claude can catch obvious issues before user reviews
- Findings become beads - Each concern is a bead with proposed fix and pros/cons
- Interactive review for details - Walk through findings one at a time
- Triage findings - Most fix on same branch, some marked as fast-follow (don't block PR)
- Flag disagreements - If implementor disagrees with a finding, surface it to user for final call
- Same session, context available - Auto-continue to next phase
- New session needed - Provide lightweight prompt with mode and epic reference
- Context lives in beads - The prompt just says what mode to enter, beads have the details
- Triggers for new session:
- Context is getting long/compacted
- User explicitly asks
- Work pattern needs fresh orchestrator (parallel subagents)
Continue implementing beads under epic beads-xxx. Enter orchestrator mode -- spin up subagents for ready tasks.
The new session runs bd show <epic> for design context and bd ready for available work.