Don't prompt it. Show it a failing test.

Prompt "add discount codes" and the agent writes eighty lines; the only way to know if it handles a discount bigger than the total is to read all eighty. Flip it: write the question as a test first, total never below zero, and prompt "make it pass". The agent tries seven times, you read one test. A month later another agent's refactor drops a check and the same test goes red in thirty seconds. A test is a prompt that never expires.

Comment below: Do you write the test or the prompt first?

Brakes: a series on the checks that let an AI coding agent go fast without shipping bugs. Every best practice shortens the time between writing a bug and finding it. Brakes are why race cars go fast. All code is real TypeScript on Node 24, typechecked and tested before it went on screen.