AI speeds up typing code but does not replace review and tests. A simple loop: task → context → diff → verify → merge.
Task format for agents
- User story + AC + files in scope + anti-goals
- Explicit stack (Nest, Nuxt, Postgres)
- Similar code example from repo
Context
@file, .cursor rules, skills — not “build the whole app in one prompt”.
Review AI code
- Verify imports and API versions
- Security: auth, validation, secrets
- Limit diff size — split tasks
Checklist
- No commit without human review
- CI mandatory
- Document AI-heavy modules
Minimal rules for solo use
- One task = one bounded context, diff < 400 lines
- Prompt: AC + files in scope + anti-goals
- Before commit: lint, test, manual scenario
- Never merge AI code you do not understand line by line