Ponytail: The Claude Code Skill That Cuts Code Output by Up to 94%

A new GitHub repo called Ponytail claims it can make Claude Code roughly 75% cheaper by forcing the agent to write far less code. The pitch is simple: AI coding agents are verbose, verbosity burns tokens, and tokens cost money. Ponytail is a single skill that makes the agent stop and ask whether code even needs to exist before it writes anything. The video walks through the repo's README, its benchmark chart, and the six-question rule that drives the whole thing.

What Ponytail Is

Ponytail is an MIT-licensed repo (DietrichGebert/ponytail), at release v4.6.0 in the frames, with around 15,000 stars per the video. The README markets it through a character: "You know him. 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 "He says nothing. He writes one line. It works." The README also notes it "works with 13 agents," so it is not limited to Claude Code alone.

The core claim, printed directly in the README: 80-94% less code, 47-77% cheaper, and 3-6x faster than a no-skill agent, measured as the median of 10 runs across Haiku, Sonnet, and Opus.

The Six-Question Ladder

The mechanism is a decision ladder. Before writing code, the agent stops at the first rung that holds:

1. Does this need to exist? -- no: skip it (YAGNI)
2. Stdlib does it? -- use it
3. Native platform feature? -- use it
4. Installed dependency? -- use it
5. One line? -- one line
6. Only then: the minimum that works

In plain terms: does the code need to exist at all, does the standard library already do it, is it a native platform feature, is there an already-installed dependency, can it be one line, and only if all of those fail does the agent write the minimum amount of code. The README frames this as "Lazy, not negligent," and notes that trust-boundary validation and data-loss handling are still kept. So it is meant to cut bloat, not safety.

The Benchmark Numbers

The benchmark chart measures median lines of code across three "arms": baseline with no skill (gray), the older caveman skill (orange), and ponytail (green). Lower is leaner. The test uses five everyday tasks -- email validator, debounce, CSV sum, countdown timer, rate limiter -- with 10 runs per cell.

Model	Baseline (no skill)	Caveman	Ponytail
Haiku	518	116	39
Sonnet	693	120	44
Opus	256	67	51

The video calls out Opus specifically: 256 lines down to 51 using the skill. Across all three models, ponytail produces the fewest lines, and since lines of code roughly track token usage, fewer lines means lower spend.

Reproduce It Yourself

The repo does not just assert the numbers. The README gives the command to run the eval yourself: npx promptfoo eval -c benchmarks/promptfooconfig.yaml. Method and raw numbers live in the benchmarks/ directory, and write-ups for production-grade tasks (where an unconstrained agent bloats much more) are in benchmarks/results/. Every shortcut ponytail takes is marked in the generated code with a ponytail: comment naming its upgrade path, so you can see exactly where it cut a corner and how to undo it.

Why It Matters

The creator frames the token savings around model cost. With Opus already expensive and Fable expected to return at a high price (and possibly off the max subscription plan), cutting code output by 80% or more directly reduces spend on every request. The argument is that a habit good engineers already practice -- delete code, reach for the standard library, avoid reinventing what already exists -- can be encoded as a skill and applied automatically by the agent. The creator says a full deep dive on the repo is coming on his YouTube.

Key Takeaways

• Ponytail is a single MIT-licensed Claude Code skill (DietrichGebert/ponytail, v4.6.0, ~15k stars) that forces the agent to minimize code before writing.
• It works by a six-question ladder, stopping at the first rung that holds: skip it, use stdlib, use a native feature, use an installed dependency, make it one line, or write the minimum.
• README claims: 80-94% less code, 47-77% cheaper, and 3-6x faster, measured as the median of 10 runs across Haiku, Sonnet, and Opus.
• Benchmark on five everyday tasks shows Opus dropping from 256 baseline lines to 51, Sonnet from 693 to 44, and Haiku from 518 to 39.
• It beats the older "caveman" skill on every model, and the numbers are reproducible via npx promptfoo eval -c benchmarks/promptfooconfig.yaml.
• It aims to be "lazy, not negligent" -- trust-boundary validation and data-loss handling stay, and each shortcut is tagged with a ponytail: comment naming its upgrade path.