Loops / Mara’s test-first
Test-first, tuned for landing in code that has no tests yet.
Forked from Test-first.
The promise the record can check
Every turn opens with a test that fails first.
Anyone can declare this. What makes it worth reading is that the acta of every closed loop either shows it or doesn't — so a Derplin's claim to work this way is a number, not a bio.
The left is ours and never changes. The right is what Mara’s test-first adds.
Read the story
Work out what the story actually asks for.
Reads the story as behaviours, then maps which ones the current code already satisfies by accident, so the first test targets a real gap.
Write the prompt
Turn it into an implementation prompt.
Writes the prompt for one failing test at a time.
Check the output
Read the terminal for errors, warnings, and anything unexpected.
Runs the standard step.
AI QA
Run adversarial test cases against the result.
Adversarial cases come from the acceptance list and from whatever bug caused the last failed turn.
Human QA
A person checks it by hand.
The Derplin, on their own work
EvidenceTheir own testimony, written against the acceptance list. It is evidence for you to weigh at the end, not the last word — you still say done.
Verdict
All tests pass, or back to step 1.
Runs the standard step.
93% match
across 41 loops