Ponytail skill for Claude Code
Ponytail is a one-prompt skill by Dietrich Gebert that makes a coding agent climb a ladder before writing anything: does this need to exist, is it already in the codebase, does the standard library or the platform already do it.
/plugin install ponytail@ponytailCoding agents over-build. Ask for a date picker and you get a new dependency, a wrapper component, a stylesheet and a question about timezones. Ponytail is a single prompt, skills/ponytail/SKILL.md, that stops at the smallest answer that holds. In the README's example, the whole date picker becomes <input type="date">. The license is MIT, and the repository passed 150,000 GitHub stars within four months of launch. For the basics, see Claude Code skills.
The ladder
After reading the code the change touches, the agent stops at the first rung that works:
- Does this need to exist? If not, skip it.
- Is it already in this codebase? Reuse it.
- Does the standard library do it?
- Does the platform have a native feature for it?
- Does an installed dependency do it?
- Is it one line?
- Only then, the minimum that works.
The README is explicit about what is never cut: validation at trust boundaries, handling of data loss, security and accessibility. The ladder trims what the agent builds, not how carefully it reads.
Commands
| Command | Does |
|---|---|
/ponytail lite | full | ultra | off | Sets the intensity or turns it off |
/ponytail-review | Reviews the current diff for over-engineering and returns a list of things to delete |
/ponytail-audit | The same check across the whole repository |
/ponytail-debt | Collects the shortcuts marked ponytail: into a list, so deferred work gets done |
/ponytail-gain | Shows the benchmark numbers |
What the benchmark says
The author ran headless Claude Code sessions on twelve feature tickets in a real FastAPI and React template, with and without the skill (Haiku 4.5, four runs each). Against no skill, Ponytail produced 54% fewer lines of code, used 22% fewer tokens, cost 20% less and finished 27% faster, with every result judged safe. The biggest cuts came where a native control replaces a component: the date picker went from 404 lines to 23. On tasks with no over-build trap the difference was close to zero. These are the project's own numbers on one repository and one model.
How to install it
Claude Code, as two separate prompts:
/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytailCodex:
codex plugin marketplace add DietrichGebert/ponytail
codex plugin add ponytail@ponytailThen trust its two hooks under /hooks and start a new thread. For any other agent, copy the repo's AGENTS.md into your project. The author warns that the only genuine sources are the GitHub repo and the @dietrichgebert/ponytail npm package, and that the project never ships .exe or .dll files.
When to use it
- Front-end work, where native inputs, dialogs and CSS often replace whole libraries.
- Reviews: run
/ponytail-reviewon a branch an agent wrote before you merge it. - Alongside Caveman. The README recommends the pair: Caveman shortens what the agent says, Ponytail shortens what it builds.
It is the wrong tool when you have decided on a heavier design on purpose. Insist and the agent will build it, but expect pushback first. For a full plan-test-review method, Superpowers fits better.
Sources
- DietrichGebert/ponytail README — GitHub, accessed October 2026