Prompting and review

Prompting

Scoped prompts that name real files produce changes small enough to review:

• Point it at the docs. Each car workspace’s docs/ folder is the source of truth, so start with “read docs/architecture.md before proposing anything”.
• Ask for a plan first. “Give me a plan; do not edit files yet.” Review the plan before letting it change anything.
• One task per session. “Fix the steering offset in vesc.yaml and update the comment” reviews well. “Improve the driver” does not.
• Ask for the diff and the test run. “Show the exact changes, then run the package tests and paste the output.”
• Name files and commands from the repository, and the build and test commands from Build and test.

Review

An AI-written change is a draft, held to the same bar as any team member’s code, because it runs on a car that can hurt someone:

1. Read every change before it runs. A change you cannot explain does not run.
2. Build and run the tests yourself. A tool reporting that the tests pass is not a test run.
3. Test new driving code in order: static checks, then wheels off the ground, then low speed on the floor.
4. Reject any change that disables the LB deadman or sets enable_deadman: false, whatever reason the tool gives.

An AI-written change merges, or runs on a car, only after a human review and a real test pass.