Does writing less code cost you quality?

Ponytail keeps its core promise: roughly half the code, for roughly half the generation cost.

Lower bars are better here. Ponytail (terracotta) is shorter almost everywhere — most dramatically on larger UI and algorithm tasks. The one exception is the trivial health-check API, where ponytail's rule of "always leave a runnable check behind" added a second test file.

The trade is real: leaner code scored about one point lower, and the gap is widest exactly where it matters.

Two independent judges scored every pair blind (folders relabelled A/B) on a 1–10 scale across seven dimensions. Scores below are averaged across both judges and all 12 tasks. Readability is nearly a tie — ponytail's code is just as clear, there's simply less of it — but robustness and production-readiness fall off sharply.

Where ponytail wins, and where it shouldn't be trusted unsupervised.

The skill's value is not uniform. On frontend work it strips the most code with the smallest quality hit. On data/CLI tasks it cut the most corners — missing edge cases that cost real points. On the pure refactor task, both agents converged on the same one-line answer, so the skill made almost no difference.

Same brief, two philosophies — rendered.

The three frontend tasks produced real, openable pages. These are live screenshots of each build. Notice how close they often are: ponytail's versions are leaner under the hood yet still look professional — the differences are in polish and edge cases, not basic competence.

All 12 builds — open any one for scores, bugs, judge notes, code and screenshots.

Ordered simple → complex. Each row un-blinds the A/B comparison, shows the objective metrics, the bugs each arm's judges actually found, one judge's head-to-head summary, and the full source of both versions.

A sharp tool, pointed at the right job.