I swear by Yegge's Beads https://github.com/steveyegge/beads. Just love it. One of the things Steve talks about in his best practice guide is that the only things that belong in Beads are things that are being actively worked on; it's not a backlog.

I also really like Claude's planning mode, except for the fact that it's tightly coupled with execution.

A pattern I've found is to work in different loops: a planning loop and an execution loop.

Planning loop: 1. Use GitHub issues for backlog items. Whenever I get an idea, I just make an issue with minimal detail. 2. Pull issue into Claude Code 3. Create branch 4. Use a home-grown planning skill that reads the Issue and operates pretty much the same way as Claude's but without the straight-to-execution annoyance 5. Once the plan is finalized, decompose it into Beads 6. Check in branch and tag as "ready"

Execution loop: Watch GitHub for "ready" issues. When one is found, checkout a worktree then start building. All the beads are self-contained in the issue so there's no need for more discussion with the LLM.

I turned the planning phase into a Claude plugin that I can invoke with a slash command, /beads-planner issue 44. The execution loop can be ralph or even just interactive CC sessions.

Everyone is coming up with their own way to build but figured I'd share because this works for me. YMMV