links: repo | project page
ai blueprint started as a way to make ai-assisted coding less slippery. this week it moved a level down: not just a Claude Code workflow, but a portable workflow layer that can live in different agent surfaces.
the practical change was adding Codex adapter skills alongside the Claude ones. that forced a useful cleanup. if the workflow only makes sense in one tool, it is probably leaning too hard on that tool’s habits. the core loop should be boring enough to survive the surface changing:
- onboard the repo
- write the plans
- generate the overview
- pick one feature
- implement one step
- check the result
- mark the step complete
i also added /doctor, which is the command i kept wanting and not having. it
does not build anything. it reads the setup and tells me whether the blueprint
overlay, adapters, commands, plans, overview, and git state look sane.
that matters because the workflow is becoming less about prompts and more about project hygiene. a good agent workflow should tell me when the project is ready to use it, when it has drifted, and what the next action should be.
the funny part is that making it support more tools made the whole thing simpler. the abstractions that survived were the ones that belonged in files, not in any one chat runtime.