Plan, write and test a code change with AI

• Pick a change you could explain to a colleague. Unclear jobs give unclear code.
• Work on a branch, never directly on the live system.
• Know your organisation's rules on which code may go into which tool. Never give an agent production keys.

Replace the words in square brackets with your own. Stay in the same chat for the next prompt, unless a step says otherwise.

1. You
   Write what should change, how you'll know it works, and what must not change.
2. AI
   Ask for a plan before any code, and correct it.

   PromptCopy

   Before writing any code, read the relevant files and give me a plan for this change: [change].
   
   It is done when: [acceptance test]. It must not change: [constraints].
   
   List the files you would touch and why, the risks, and anything you are unsure of. Ask me up to three questions if the job is unclear. Do not start until I agree.

3. You
   Fix the plan. If you can't judge it, the change is too big or too unfamiliar: split it.
4. AI
   Write a failing test first, then the smallest change that passes it.

   PromptCopy

   Write a test that fails now and will pass when the change is done. Run it, or give me the command if you can't. Show me the failure, and confirm it fails for the reason we expect. Then make the smallest change that passes it. Do not touch unrelated code, do not add libraries without asking, and do not change or delete existing tests to make them pass. If an existing test conflicts with this change, stop, show me the test and why, and wait. Do not skip it.

5. You
   Read the whole diff line by line, and run the full test suite yourself.
6. You
   Get a person to review it as usual, and tell them what the AI wrote.

• You can explain every line. If you can't, it doesn't merge.
• Tests weren't weakened, skipped or deleted to get a pass.
• Any new library really exists and is the one you meant. Made-up package names can be taken over by attackers.
• Look at input handling, permissions and secrets: an AI's code can run correctly and still be unsafe.

• Payment, safety-critical or security code, unless your normal extra review applies in full.
• A large change to a system you don't know. Use AI to explain the code first instead.
• You are timing yourself against a deadline and the review would be skipped.