tdd

The agent codes better with a tight feedback loop than with a long specification. A failing test is the tightest loop there is: it states the target, and the target either goes green or it does not.

The loop

1. Red. Write one failing test that states the next slice of behavior. Run it. Confirm it fails for the right reason - a test that passes before you write the code is testing nothing. If it errors instead of failing on the assertion, fix the test setup first.
2. Green. Write the least code that makes the test pass. Not the general solution, not the abstraction - the minimal move. Run the test. Green.
3. Refactor. Now clean up with the test as a net: remove duplication, name things, simplify. Re-run after every edit. Behavior is frozen; only the shape changes.
4. Repeat for the next slice. Small slices. The rate of feedback is the speed limit - never take on a step too big to hold a single test.