Keep people in the loop: review AI proposals and decide what becomes canonical.
Validation proves a change is correct; human review proves it is right. AI proposes changes, but a person reviews the diff and decides whether to accept, edit, or reject before it is committed.
The acceptance gate
Review the diff
Read the proposed model and spec changes as a branch diff.
Judge intent
Confirm the change matches the design intent, not just the schema.
Edit or reject
Adjust details directly, or send it back for another pass.
Accept and commit
Commit the reviewed change so it becomes part of the branch history.
What reviewers check that validation cannot
| Validation checks | Humans check |
|---|---|
| Correct structure | Correct intent |
| Resolvable references | Sensible relationships |
| Valid syntax | Clear naming and responsibilities |
| Consistency | Fit with the broader architecture |
AI proposes, humans decide
Acceptance is a human action. Treat AI output as a proposal that must earn approval, just like any teammate’s pull request.
Keeping a clean trail
- Commit AI-assisted changes with messages that note the intent.
- Use branches so AI work is isolated and reviewable.
- Prefer small, focused proposals over large generated dumps.
Review AI changes with the same rigor as human changes — no more, no less. Consistent review standards keep the model trustworthy regardless of who (or what) drafted the change.
Last updated on