Steinberger's test cleanup prompt names a 20% target
The instruction that did the work was a number, not a request to tidy up.
The prompt behind OpenClaw shedding around 400k lines of its own tests was not "clean this up". Peter Steinberger says the wording that worked was to remove 20% of the least useful tests while maintaining code coverage within 2%, and that the deletion happened without much change in code coverage. The scale of the cut was covered earlier. The part worth copying is the target.
If you just tell the agent to clean up, it will stop far too early.
Modern models just love writing tests for every tiny change, even if they aren't useful.
Steinberger's diagnosis of why the pile got so big is short. Modern models, he says, love writing tests for every tiny change, even if they aren't useful. That is not a bug you notice on any single commit. Each new test looks reasonable on its own, passes, and gets merged, and a year of that leaves you with a suite nobody would have designed on purpose.
Why the number matters
An open-ended cleanup instruction gets you a cosmetic result. Steinberger's line on it is blunt, and it is the whole lesson in one sentence.
A percentage does two useful things at once. It sets a floor the agent has to reach, so it cannot declare victory after trimming a handful of obvious duplicates, and it pairs the cut with a constraint that stops the cut from being reckless. Coverage within 2% is the guardrail. Without it, "delete 20% of the tests" is an instruction an agent can satisfy in the dumbest possible way.
The business read is that this is a prompting habit, not a testing insight. If you are shipping with agents all day, you are accumulating code you did not ask for in exactly this way, and the correction is to hand the agent a measurable goal plus a measurable limit rather than a vibe. One founder's result on one codebase is not a rule, and Steinberger's numbers are his own. But the shape of the prompt costs nothing to try on a suite you already suspect is bloated.
Steal the numbers, not the vibes
One email a day. Under three minutes. Real MRR, real pricing, real launches from founders building with AI.
