treat prompt changes like code

prompt files were living as freeform notes in chat history. that lasted until two people edited the “same” system prompt and neither knew which was live.

what we tried

one markdown file per agent role. commits for every material change. a short changelog at the top of the file: date, author, one sentence.

we also started pairing each change with at least one new or updated eval fixture.

what worked

review. a pull request on a prompt sounds formal until you watch a silent regression disappear in five minutes of diff reading.

named versions (planner-v3) make rollback boring instead of heroic.

what’s next

a tiny CLI that prints the active prompt hash into agent run logs so we can tie a failure back to the exact text that produced it.