AI in the Build Loop [yoe]

Yoe build system integrates Claude Code AI skills into the build loop, allowing developers to diagnose and fix build errors via a terminal UI or Claude session. The 'diagnose' skill reads build logs and automatically corrects issues, demonstrating how AI agents can handle complex build-system reasoning that humans find difficult.

A build system is a lot of small pieces — feeds, units, dependencies, task steps, sandboxes. None of it is hard on its own. The hard part is holding all of it in your head at once, which is exactly what humans are bad at and agents are good at. The latest video https://youtu.be/nfG UTerYeo shows how yoe uses that: it ships Claude Code skills you install into your project, and wires a diagnose step into the build loop. The AI skills page https://docs.yoebuild.org/ai-skills.html covers both paths. Run yoe skills install inside your project and it drops the skills into your .claude/ folder. yoe init now does this by default. Keeping them local lets you list, update, and diff them against your own edits. Note that update overwrites, so rename a skill first if you’ve customized it. The other path is the Claude Code plugin marketplace: add the yoe marketplace, reload, and skills like diagnose and new-unit show up as plugins. Several skills ship in the set. Many are still ideas, but diagnose and new-unit are in daily use, and the rest will fill in over time. d to diagnoseThe demo: add a bad echo make error task step to the ca-certificates unit. yoe sees the unit is no longer cached, rebuilds it, and the build fails on the first step. You could shell out and hunt for the problem yourself. Instead, press d . That runs the diagnose skill and hands it the build log, so the agent starts from the evidence instead of searching for it. It reads the log, finds the bad line, removes it from the unit, and the rebuild succeeds — all without leaving the yoe terminal UI. You can also run it from a Claude session: say “diagnose the unit that fails to build” and get the same result. The TUI path keeps you in the build loop; the Claude path keeps your agent and build side by side. Either way, the agent gets the log instead of guessing. This is a trivial example on purpose — a one-line error, a one-line fix. The point is the wiring, not the bug. The same pattern reaches harder problems as the skills mature: porting BSPs from Yocto, expanding new-unit , and pulling units across from Alpine, Debian, and other systems. The full set of walkthroughs is on the Videos page https://yoebuild.org/videos/ . A few ways to go further: /new-machine and /new-module AI skills