Experimenting with Steve Yegge's beads for large project task tracking

[ciaran-finnegan]

I think on larger projects the PRD and plan files might be getting too big. The Algorithm loads them into context and that could be a lot of context gone before any actual work starts. I suspect when compaction kicks in the agent loses its place but I'm not 100% sure that's what's happening.

I'm experimenting with Steve Yegge's beads to track tasks instead. bd ready --json gives the agent the next task without loading the whole plan into context. There's a Decompose skill that takes a PRD or implementation plan and breaks it into atomic beads issues with acceptance criteria and dependency chains — so the agent doesn't have to figure out task ordering itself.

Early days but it seems to help.

Fork: https://github.com/ciaran-finnegan/Personal_AI_Infrastructure/tree/ciaran/customizations

[DolphusCY]

@ciaran-finnegan im doing the same thing. I integrated beads with PAI yesterday.

[ciaran-finnegan]

Keen to hear how it goes for you @DolphusCY

[javierburongarcia]

Very interesting, are you using a fixed threshold to define which ones are too big, or is up to the model to go one way or the other?

[klittle32]

I look forward to hearing how your experience with beads and PAI goes. I used beads for a small project just before they migrated it to dolt and I really enjoyed it. Now that the dust has settled with regard to dolt and they ship a standalone dolt binary with beads, I think I'll use it for my next project. I haven't done any real comparisons but my impression was that Claude managed to stay on task quite well with it and context was not polluted.