Team splits reason fields on Claude Code project to label whose judgment is whose

The team's own definition of intent debt has two halves: reasoning that can't be traced later, and AI's "intent-looking" output being preserved as if it were a human decision. The first half is old — when a veteran leaves, the reasons leave with them. What AI changed, in their account, is the volume: one person now makes far more decisions per day with AI than before. The second half is newer, and they tie it to Addy Osmani's "The Intent Debt," which warns that an AI will invent a confident-sounding reason, which is worse than admitting it doesn't know.

Their answer was structural rather than behavioral. The AI's reasoning is kept, because it has value as research, but it is not allowed into the human rationale field; when a person offers no reason, the record says so explicitly rather than leaving a blank. Enforcement comes from a PR template section that a human writes, checked by a CI script that only verifies the line exists — never that the words are human, because no machine can prove that.

Where the system is honest about its gaps is instructive. The /decide skill may simply not fire, in which case no DEC ID appears and CI cannot notice. When Claude was turned loose on 98 migrated files, the token stream outside comments matched exactly and type checks and lint were unchanged, yet review found three errors — including a comment rewritten to cite a reason that was never in the original, the very failure mode the project is trying to prevent.