Ponytail, a free 'laziest senior dev' prompt for Claude Code and other coding agents, cut a 335-line date picker to 10 lines by using the browser's own calendar
Dietrich Gebert's Ponytail is one prompt file that makes a coding agent check what already exists before writing anything. His 585-session benchmark says it cut code 53% and cost 26%, and admits where it lost.
Ask a coding agent for a date picker and, in Dietrich Gebert’s benchmark, it wrote 335 lines. With his skill loaded, it wrote 10.
“Without Ponytail, the agent installs a date picker library or builds a whole calendar by hand,” the Ponytail README says. The skill’s version “first looks at what is already there: the repo has an Input component, and every browser has a date picker. It puts the two together.”
That’s the whole trick, and it’s a persona. “Long ponytail. Oval glasses. Has been at the company longer than the version control. You show him fifty lines; he looks at them, says nothing, and replaces them with one.” The tagline is shorter: “He says nothing. He writes one line. It works.”
Under the joke is one Markdown file, SKILL.md, that opens with “You are a lazy senior developer. The best code is the code never written.” It gives the agent a ladder and tells it to “take the first option that fully works”: does it need to exist, is it already in this codebase, is it in the standard library or the platform, is it in a dependency you already have, can it be one line, and only then the minimum code. A shortcut with a known limit gets a comment saying so. The README lists what’s off the table: “trust-boundary validation, data-loss handling, security, and accessibility are never on the chopping block.”
Version 5 shipped October 8, and Gebert’s benchmark notes are the interesting part. He ran 39 tasks five times each across three setups, 585 headless Claude Code sessions on Opus 5.5. Twelve were feature tickets in a real FastAPI and React template. Eighteen had hidden checks for things like SQL with untrusted input, path traversal and forged tokens. Against the same agent with no skill, Ponytail 5 wrote 53% fewer lines, ran 41% faster and cost 26% less, with mean cost per session falling from $0.200 to $0.142.
It didn’t get sloppier. The hidden checks passed 86 of 90 without the skill and 87 of 90 with it, and on six security tasks every setup passed all 30 runs. Ponytail wrote a test where the logic needed one 98% of the time, against 68% without it.
Then the benchmark lists where it lost. To grade the tests, Gebert injected small bugs into each agent’s code and ran its own tests against them. Where a test existed, the no-skill tests caught 77% of the bugs and Ponytail’s caught 69%. “No skill writes the most thorough single tests, but often writes none,” the notes say. A blind judge reading the agents’ replies leaned slightly toward the longer no-skill ones, 106 to 82, which the notes call no significant difference. The notes even own up to a harness bug: an earlier version ran some tasks inside the Ponytail repo, where Claude Code showed the model the repo’s recent commit messages.
The repo, created in June, sits at more than 174,000 stars. Version 5.3 added a band above the Claude Code prompt that shows what the current task costs and how often the agent reused code instead of writing it. Whether it holds up on a codebase nobody tuned it against is the test every user runs alone.
The README’s FAQ asks “What if I really need the 120-line cache class?” and answers: “You don’t. Insist anyway and he’ll build it. Slowly. Correctly. While looking at you.” Nobody has published how often a user insists.