Make Claude Code write less code: the Ponytail skill

If you have ever watched an AI agent write 500 lines for something that only needed 50, this is for you. The agent is not being lazy, it is being the opposite: it happily rebuilds things that already exist and pads out simple jobs. More code sounds productive, but it usually means more bugs, more to maintain, and a bigger bill.

Ponytail is a skill by DietrichGebert (over 100,000 stars on GitHub) that fixes this. It forces the agent through a short decision ladder before it writes anything.

What it actually does

Before writing new code, the agent has to answer, in order:

• Does this code need to exist at all?
• Does it already exist somewhere in the project?
• Can the standard library, a built-in platform feature, or a dependency you already have do it?
• Only if all of those are no does it write something new.

The tagline sums it up: the best code is the code you never wrote.

Install it (Claude Code)

It is two commands. Open Claude Code and run:

/plugin marketplace add DietrichGebert/ponytail

/plugin install ponytail@ponytail

That is it. From now on it runs on your coding sessions automatically.

Using another editor?

Ponytail also ships rule files for other agents. Copy the matching file from the repo into your editor's rules folder:

• Cursor: .cursor/rules/
• Windsurf: .windsurf/rules/
• Cline: .clinerules/
• Anything that reads AGENTS.md: drop it in at the project root.

What to expect (honest numbers)

On the author's own benchmark (a real full-stack repo, 12 feature tickets), Ponytail produced up to 54% less code, about 20% cheaper, and 27% faster, using around 22% fewer tokens.

Two honest caveats so you are not disappointed:

• That 54% is a best case, driven by obvious over-build traps (one date picker went from 404 lines to 23). On code that is already lean, the savings are close to zero.
• An independent JetBrains test across 80 paired tasks found smaller real-world gains, around 15% less code and 10% cheaper. Still a free win, just not magic.

The point underneath the numbers is the good part: it only writes what the task actually needs, and it never cuts corners on validation, error handling, or security to hit a smaller number.

Links

• Repo: github.com/DietrichGebert/ponytail
• Benchmarks: the author's benchmark folder
• Independent test: JetBrains