Claude Code Wrote 335 Lines. With Ponytail, It Wrote 10

Same ticket, same agent: "add a date picker to the frontend." Without help, Claude Code wrote 335 lines. With a free skill called Ponytail, it wrote 10. That's from Ponytail's own benchmark, so here's what the numbers show, what they don't, what the latest version adds, and how you could turn the tool into a small paid service.

What Ponytail is

Ponytail is an MIT-licensed skill by Dietrich Gebert. The repo describes it as making "your AI agent think like the laziest senior dev in the room": the colleague who looks at your fifty lines and replaces them with one. The whole thing is one prompt, skills/ponytail/SKILL.md, plus the files that load it into Claude Code, Codex, Cursor, Copilot and other agents.

The prompt gives the agent a ladder to climb before it writes anything, and it stops at the first rung that works:

1. Does this need to exist at all?
2. Is it already in this codebase?
3. Does the standard library or the platform already do it?
4. Is there an installed dependency for it?
5. Can it be one line?
6. Only then: the minimum code that works.

It also sets limits on the laziness. The README says it will "never cut validation, error handling, security, or accessibility," and that logic with a branch, a loop, a parser, money or security "leaves one small test behind." Every reply ends with what the agent skipped or didn't check.

The project has been around since June 2026. The current major version, Ponytail 5, shipped on Oct 8, and the latest release is v5.3.3 from Oct 29, 2026. It had about 184,000 GitHub stars when I checked on Nov 4.

New in 5.3: the band

Since v5.3 (Oct 16), Ponytail shows a band above the Claude Code prompt: what the current task costs, how many tokens and how much agent time it used, how often it reused existing code, and whether the tests passed after the last change. Its buttons switch the level, run a review or an audit, and open the board, a local page with every task across your projects. The README says everything stays on your machine by default, in ~/.config/ponytail/band/, and /ponytail band hides it.

It needs Claude Code 2.1.290 or later, in the terminal or the desktop app. It doesn't show in the VS Code chat panel, cloud sessions or claude -p. The v5.3.3 release on Oct 29 was a round of fixes for it.

The before and after

In the date picker task, the agent without the skill built a whole calendar by hand: 335 lines (the median of 5 runs). With Ponytail 5, it noticed the repo already had an Input component and that every browser ships a date picker. It put the two together:

function DatePicker(props: Omit<React.ComponentProps<typeof Input>, "type">) {
  return <Input type="date" data-slot="date-picker" {...props} />
}

That's the core of the 10-line file. The median cost of that task dropped from $0.43 to $0.10 at list price.

The numbers, and who measured them

The author ran the benchmark on Oct 7, 2026, on Ponytail 5 (before the band and the later fixes): 39 tasks in headless Claude Code sessions with Opus 5.5, 5 runs each, comparing the same agent with no skill, the previous Ponytail and Ponytail 5. Twelve of the tasks were feature tickets in a FastAPI + React template repo.

Against no skill, on a typical task (the geometric mean of each task's median), Ponytail 5 measured:

• 53% fewer lines of code
• 26% lower cost
• 41% less time
• 45% fewer output tokens

The date picker is one of the biggest gaps. Some tasks came out the same size (a caching task was 11 lines in every version), and a few came out a little bigger.

The catch

It's the author's own benchmark. The write-up is in the first person, by the person who wrote the skill. The method, the per-task tables and the limits are all published, and the harness ships in the repo so you can rerun it, which is more than most tools offer. Still, treat the percentages as the author's measurement.

Less code didn't mean more correct code. Eighteen of the tasks had hidden checks that ran the code. No skill passed 86 of 90 runs and Ponytail 5 passed 87 of 90. The author says it plainly: less code "did not cost correctness or safety. It did not add any either."

Two smaller notes. Bash was disabled for every agent in the benchmark, so it measures what the agent writes, not how well it debugs. And the README warns to install only from DietrichGebert/ponytail on GitHub or @dietrichgebert/ponytail on npm.

Try it in 5 minutes

1. In Claude Code, run /plugin marketplace add DietrichGebert/ponytail, then /plugin install ponytail@ponytail as a second prompt.
2. Start a new session and give it a ticket you'd normally hand an agent.
3. Compare the diff with what you usually get. /ponytail off switches it off for a side-by-side.
4. Run /ponytail-review on your current changes, or /ponytail-audit on a whole repo.

Here's how you get paid

My opinion, and no income is guaranteed: I'd bet plenty of teams now have repos that AI agents grew fast, with more code than the job needed. A short, fixed-scope cleanup is an easy first engagement to pitch.

Sell a cleanup sprint on one repo:

• The audit. Run /ponytail-audit. The README says it checks the whole repo, then "ranks what it finds and tells you what to fix first," from must fix down to nice to have.
• The fixes. Fix the must-fix items first, and review each change yourself (or with /ponytail-review) before it merges. You're accountable for the result, so read every diff.
• The handover. Give the team the ranked list, what you fixed and what's left.

The audit gets the conversation started. The fixes are the paid work.