Repository AI Workflow Policy
Use this as a baseline team policy for AI-assisted coding.
Approved Use
- Tests.
- Documentation.
- Small feature work.
- Bug fixes with reproducible failures.
- Mechanical refactors.
- Local explanations of existing code.
Restricted Use
Choose and document risk-proportionate review and authority for these areas; the list is not a universal ban on authorised investigation:
- Authentication or authorization.
- Payment, billing, or financial logic.
- Database migrations.
- Security-sensitive infrastructure.
- Large rewrites.
- Customer data handling.
- Regulated systems.
Task Brief Required
Every AI-assisted code task should include:
- Goal.
- Relevant files/modules.
- Expected behavior.
- Non-goals.
- Edge cases.
- Test command.
- Security/data constraints.
- Existing pattern to follow.
Hard Rules
- Do not paste secrets into AI tools.
- Do not use personal AI accounts for private company repositories.
- Follow the repository’s dependency policy and existing authorised scope.
- Do not bypass tests, linting, typechecking, or review.
- Require an inspectable architecture rationale and the configured review for large rewrites.
- Do not merge AI-generated changes without code review.
Review Checklist
- Follows local architecture.
- Preserves public behavior unless intentionally changed.
- Does not weaken auth, validation, logging, accessibility, or error handling.
- Tests cover behavior and failure paths.
- No secret, credential, or private customer data exposure.
- PR description states what changed, what checks actually ran and what remains unverified; follow applicable disclosure policy.
Exact-revision handoff
- Task, repository, branch/worktree and scope:
- Allowed data and enforced tool/network permissions:
- External-action/merge/deployment authority:
- Candidate commit:
- Behavioural test commands and results:
- Independent review decision for that commit:
- Material edits after review and renewed checks:
Illustrative search fixture: a term present only in another locale must not appear in the English results. Verify expected public behaviour rather than reproducing the filter expression in the test.