Ponytail Has 154,000 Stars for a Seven-Step Ruleset Whose Benchmark Is Four Runs on One Repo
Software / news
Ponytail Has 154,000 Stars for a Seven-Step Ruleset Whose Benchmark Is Four Runs on One Repo
The MIT-licensed rule file tells coding agents to write less code. Its 54% cut in lines comes from the author's own test with Claude Haiku 4.5.
Ponytail, a rule file that tells coding agents to write as little code as possible, gained 1,894 stars in a day on GitHub trending on October 4, 2026, and sits at 154,408 stars. Its published evidence is a benchmark of twelve feature tasks on one FastAPI and React repository, run with Claude Haiku 4.5, with n=4.
The repository, DietrichGebert/ponytail, is owned by an individual account, carries an MIT licence and was created on June 12, 2026, according to GitHub's API. The same API lists 227 open issues. The README page lists 55, so the count depends on where you read it.
The seven rungs a model is told to climb
Ponytail is not code. It is a ruleset that an agent reads before writing, a decision ladder applied after it understands the problem:
- Does this need to exist? If not, skip it.
- Is it already in the codebase? Reuse it.
- Does the standard library do it? Use that.
- Is there a native platform feature? Use it.
- Does an installed dependency do it? Use it.
- Does it fit in one line? Write one line.
- Otherwise, write the minimum that works.
The README says "validation, data-loss handling, security, and accessibility are never on the chopping block." It lists plugins or instruction files for more than 20 agents, among them Claude Code, Cursor, GitHub Copilot, Codex and Gemini CLI.
What the benchmark measured
The README reports reductions against a no-rule baseline, from "real agent sessions" on the twelve tasks. Every figure is the author's own.
| Metric | Reduction with ponytail |
|---|---|
| Lines of code | 54% (up to 94%) |
| Time | 27% |
| Tokens | 22% |
| Cost | 20% |
- Lines of code54 %
- Time27 %
- Tokens22 %
- Cost20 %
Source: ponytail README, author-supplied, accessed 2026-10-04
The README's best example is a date picker that fell from 404 lines to 23. It says the gains are largest where an agent over-engineers and small where the code is already tight.
It also claims "100% safe" and says ponytail is "the only arm that cuts every metric, and the only one that stays fully safe while doing it." The word arm implies other conditions were compared, and the page does not say what they were or how safety was scored.
What four runs cannot show
Four runs per task on a single repository is a small sample. Haiku 4.5 is a small model, and an instruction to write less may behave differently on a larger one. Counting lines rewards terse code, not code that is easier to maintain. A 23-line date picker is shorter than a 404-line one, and nothing in the README shows the two do the same job.
Stars are not adoption either. A repository gaining 1,894 in a day tells you how many people clicked, not how many kept the rule file installed after the first week.
The cost saving is the figure to check against your own bill. Whether a 20% drop survives on a repository with a different shape is untested.
Where it sits among agent skill packs
Ponytail belongs to a growing shelf of rule and skill packs that spread by star count. We have also covered Agent-Reach and Dwarfstar, two other repositories that collected stars quickly.
The practical test is cheap. Install the rule on a branch, give the agent two tasks you have already merged, and compare diffs and token counts. The author's repository is the only source of the 54% figure until someone else publishes a reproduction.
Sources
More in Software
- 01Levien's Next-100 Rust GUIs Post Counts About 20 New Toolkits and Bets Against egui's Immediate ModeThe Linebender author credits AI rewrites for the surge; egui's own README concedes the trade-off he is pressing on.
- 02C++26 Is Reopened: The 13–9–12 Ballot, the Enforced Label That Answers It, and What Is Still OpenSutter's Búzios report says the committee kept contracts, added one paper, and will send C++26 to a second ballot with no published date.
- 03OpenAI Says Codex Cannot Run in Docker for iOS Builds, and Docker Launches It With Codex's Sandbox OffBolin's November 18 post lists Seatbelt, Bubblewrap and MXC as Codex's three sandboxes; Docker's page defaults to --dangerously-bypass-approvals-and-sandbox.
- 04CLI-Anything Has 62,119 Stars and a README With Three Different Test TotalsThe HKUDS project turns desktop software into agent-callable CLIs; its own per-harness lines sum to 158 more tests than it reports.