Ponytail makes your coding agent stop over-building

Ask an agent for a date picker and you get a dependency, a wrapper component, a stylesheet and a conversation about timezones. Ponytail is an MIT-licensed skill whose entire job is to make the agent answer with <input type="date"> instead.

What it is

A prompt-level persona and rule set — the laziest senior developer in the room — installed into Claude Code, Cursor or a similar agent. The ladder it enforces is the one experienced engineers use without naming it: does this need to exist, is it already in the codebase, does the standard library do it, does a native platform feature cover it, can it be one line, and only then write something.

At 103,169 stars it is the most-starred project covered on this site so far.

Why it showed up now

v4.9.0 in early August, and a set of benchmark numbers that got revised downward in public. That second part is the reason it is here.

How it actually works

The skill is instructions, not code, and the interesting engineering is in the measurement. The current headline — roughly 54% less code, up to 94%, about 20% cheaper and 27% faster — comes from a headless Claude Code session editing a real repository, tiangolo's full-stack FastAPI template, scored on the git diff it leaves behind. Twelve feature tickets, the same agent with and without the skill, n=4, on Haiku 4.5.

The cut is largest where there is a genuine over-build trap: a date picker going from 404 lines to 23, a colour picker from 287 to 23, because the agent reaches for a native input instead of building a component. On code that was already minimal, the difference is near zero.

Earlier versions of the README advertised 80–94% less code as a flat figure, from a single-shot benchmark comparing bare model completions. An issue on the repository pointed out that a chatty baseline pads its answer with prose and alternatives, so much of that gap was an artefact of the comparison. The project kept the old numbers visible, marked them as superseded, and published the corrected agentic measurement as the defensible one.

The safety claim is measured too: the README states the skill keeps every safety guard, where a bare "write one-liners" prompt drops one.

Try it

Install it as a skill in your agent from the repository, then give it a task with an obvious over-build trap — a date input, a colour picker, a modal — and read the diff.

Where it is weak

The measurement is still the author's own, on one repository, one model and four runs per task. n=4 on Haiku 4.5 is a signal, not a finding, and lines of diff is a proxy that rewards terseness even where terseness is not the goal.

136 issues are open, and the last push was August 7 — a week before the snapshot, which for a project at this level of attention is a noticeable pause.

The deeper caveat is what a laziness bias costs. Most tasks benefit; some do not. Input validation, error handling and abstractions you genuinely will need in two weeks are the cases where "the shortest diff that works" is the wrong instinct, and no prompt reliably knows which case it is in.