Getting Claude Code to write the failing test first, it always writes the code too

Make "stop" a concrete, checkable step instead of a wish:

Write a failing test for X in tests/x.test.ts. Run it. Show me the failure output. Then stop and wait for me. Do not edit anything outside tests/.

The "show me the failure" part matters. If it can't produce red output, the test isn't testing anything new, and you learn that before any code exists.

Then a second message: "Now make that test pass with the smallest change. Don't modify the test."

Two messages, not one. Asking for both in one prompt is where it collapses them.

Structural fix for the sneaking: a hook that blocks edits outside tests/ during the red phase, or simply check git status after step one and throw away anything in src/.

Also look at what it asserts. A test written by the same model that will write the code tends to test the shape it already has in mind. I write the test names myself ("rejects a booking in the past", "allows two bookings back to back") and let it fill in the bodies. That's where the value is.

Plan mode helps too: ask for the test in plan mode, approve, it writes only what the plan said. Not bulletproof, but less sneaking.