He writes 94% less code, 20% cheaper and 27% faster.. and it works!

This month, a single-author repo called ponytail collected tens of thousands of GitHub stars in a couple of weeks on one promise: install a few hundred words of instructions and your coding agent stops over-building.

Its original pitch was a single flat number, roughly 80–94% less code.

Then a contributor pointed out that the number was measured against a bare chat model that pads its answers with prose, caveats, and three alternative implementations, and the maintainer rebuilt the whole benchmark!

The corrected figure, measured against a coding agent actually doing the job properly, was about 54% on average (up to 94%), enormous where the agent was about to over-build, near zero where the code was already tight.

The coding agent’s reflex to over-build is a measurable artifact of how it was trained and judged, the lazy skills are a runtime correction for that artifact, and the same length bias that bloats the code also corrupts the benchmarks meant to prove the correction works, so the interesting engineering is measure the constraint without letting the bias score its own homework.

In this article, I will skip the character study (yes, the senior dev with the ponytail and you have met him) and go straight to four things: