Ponytail is a free, MIT-licensed skill that tells Claude Code and Codex to write less code. Its author's benchmark shows 53% less code and 45% fewer output tokens, but cost fell 26%, not half. Here's what the numbers measure, and why Australian dev teams should install it, and any agent skill, only from the official source.
Ponytail is a free, MIT-licensed skill that tells Claude Code and Codex to write less code. Its author's benchmark shows 53% less code and 45% fewer output tokens, but cost fell 26%, not half. Here's what the numbers measure, and why Australian dev teams should install it, and any agent skill, only from the official source.
A short video has been doing the rounds claiming a text file on GitHub "halves Claude's tokens". The project is real. It's called Ponytail, it has more than 182,000 GitHub stars, and the idea behind it is a good one. The video still gets two things wrong: what the numbers say, and where you should get it from. The second one matters more.
We're an AI agency in Sydney, and we spend a lot of time inside coding agents. Here's our read.
Ponytail is a prompt that makes an AI coding agent behave like the quiet senior developer who replaces your fifty lines with one. The repo is DietrichGebert/ponytail, published by the GitHub account DietrichGebert under the MIT licence. It was created on 12 June 2026, and the latest release is v5.3.3, published on 29 October 2026.
The core rule, in the README's words: "write only what the task needs, and never cut validation, error handling, security, or accessibility."
Before writing anything, the agent walks a ladder and stops at the first rung that holds:
The README's own example: asked for a date picker, the agent without the skill wrote 335 lines. With Ponytail it wrote a 10-line file that reused the repo's existing input component and let the browser supply the calendar.
That's a sound engineering instinct. Most of us have inherited a codebase full of hand-rolled things the platform already did.
The README says "Ponytail is one prompt", and that's true of the core: the SKILL.md file. But what you install depends on your agent. The Claude Code and Codex plugins also run two small Node.js lifecycle hooks, and Codex asks you to review and trust them. The Claude Code version adds slash commands, a status band showing cost and tokens, and a local board that stores session history on your machine. The repo holds plugins and rule files for a long list of agents, including Cursor, Copilot, OpenCode and Gemini.
None of that is sinister. It does mean "it's only a prompt" is the wrong mental model when you decide whether to trust it.
The headline on the README is: -53% code, -41% time, -26% cost, -45% tokens. These are the author's own numbers, from a benchmark file dated 7 October 2026 that he published in the repo: benchmarks/results/2026-10-07-agentic.md.
To the author's credit, the method is written down in detail. In short:
The numbers are medians per task, compared with the no-skill baseline, then averaged across tasks, with 95% intervals. The -53% code figure, for example, has an interval of -63% to -41%.
So it's a careful benchmark. It's also small, self-reported and not independently replicated, on one model in one harness. Your results on a different agent, model or codebase could be quite different.
This is the gap the video skipped. The 45% figure is output tokens, meaning the text the model writes back. Cost is the price of the whole session. In the benchmark file, average cost per session went from $0.200 with no skill to $0.142 with Ponytail 5, at list price.
The benchmark file doesn't explain why cost fell by less than output tokens, so we won't guess. The practical point is simpler: if someone tells you it halves your bill, that's not what the author measured. A 26% saving on these tasks is still worth having. It isn't half.
The video makes a point of Ponytail passing all 30 runs on six security tasks: SQL injection, path traversal, forged tokens, rate limiting, malformed CSV rows and caching. True. But the benchmark file says all three arms passed 30 of 30, including the agent with no skill at all.
The author is clear about this: "Less code did not cost correctness or safety. It did not add any either." On the 18 tasks with hidden checks, no skill passed 86 of 90 and Ponytail 5 passed 87 of 90. That's level.
So the fair reading is that, on this benchmark, writing less code didn't make the output less safe. It's not evidence that Ponytail makes code more secure.
Where Ponytail 5 stands out, per its author, is tests. On tasks where the logic needed a test, Ponytail 5 left one behind 98% of the time, against 68% without the skill. The agent's own tests caught 66% of deliberately injected bugs, against 46%.
There's nuance here too. Where a test existed, the no-skill tests caught more bugs (77% against 69%). The author's summary: "No skill writes the most thorough single tests, but often writes none."
Ponytail's rule is the right default for most business software. Less code means less to review, less to maintain and less surface for bugs. That matches how we want agents to work on client projects.
But we'd treat it the way we'd treat any style guide for a junior developer. It nudges behaviour. It doesn't replace code review, tests or a security pass. And the savings are worth measuring on your own work before you put them in a business case.
If your team wants to try it, a two-week trial on a real repo, tracking cost per task and review time, tells you more than any README.
Here's the part that matters more than any benchmark.
The video ends by sending viewers to the creator's Telegram channel for "the link". For a free, public GitHub project, that's an unnecessary detour, and this project has already had a lookalike problem.
In August 2026, a user reported in the project's issue tracker (issue #735) that a trojanized clone of Ponytail had been distributed as a Claude Code plugin marketplace and shipped Windows malware. The clone was a separate repo under a different account, not the official project. The maintainer closed the issue on 5 October, noting that the clone's repo and account were gone, and added a line to the README naming the official sources:
The README also says the project never ships .exe or .dll files, and that a copy that does isn't his. We're deliberately not repeating how the clone worked. If you think you installed something from a lookalike, start with the issue itself and your endpoint security tools.
The lesson isn't really about Ponytail. Popular agent skills are now worth copying, and a copy one letter off the original looks the same in a terminal. The ASD Cyber Threat Report made a related point this year: AI is shrinking the time between a weakness going public and someone using it. Trust in a popular name is exactly what a lookalike borrows.
This is the hygiene we'd want any Australian dev team to follow before enabling a Claude Code, Codex or Cursor skill, Ponytail or otherwise.
A skill is instructions your agent will follow, and sometimes code it will run. Before you turn one on:
If you only want the rules and not the extras, Ponytail offers an instruction-only path: copy its AGENTS.md or rule file into your project. You lose the commands and mode switching, but nothing executes. That's a reasonable lower-trust way to trial it.
The ASD's Essential Eight wasn't written for AI agents, but application control and restricting admin privileges apply directly. For tools that check skills and agents themselves, see our roundup of open source AI security tools for AI agents.
Picture a hypothetical eight-person software studio in Surry Hills. A developer sees the video and wants Ponytail. Under a secure-by-default policy, they install it from DietrichGebert/ponytail, a second developer reads the SKILL.md and the two hooks, the version goes in the team's tooling doc, and the agent runs in a dev container without production keys. Twenty minutes of work. If a lookalike turns up next month, it doesn't get through the door, because nobody installs from a link.
If your code or client data can't leave Australia, the agent setup matters as much as the skill. Our private AI in Australia page explains how we approach that.
Ponytail is a free, MIT-licensed skill and plugin that tells AI coding agents such as Claude Code and Codex to write only the code a task needs, reuse what already exists, and never cut validation, error handling, security or accessibility.
No. In the author's own benchmark, output tokens fell 45% but cost fell 26%. The benchmark used 39 tasks on Opus 5.5 in Claude Code, and it hasn't been independently replicated.
The benchmark doesn't show that. On its six security tasks, the agent with no skill also passed all 30 runs. The author's conclusion is that less code didn't cost safety, and didn't add any either.
The author names two official sources: DietrichGebert/ponytail on GitHub and @dietrichgebert/ponytail on npm. A trojanized lookalike circulated in August 2026, so avoid copies from links in videos, DMs or chat channels.
Verify the publisher, pin the version, read the prompt file and any hooks or scripts before enabling it, look for unexpected binaries, and run the agent with least privilege, away from production credentials.
It's released under the MIT licence, which allows commercial use. Your agent and model usage are billed separately by their providers.
We help Australian businesses put AI agents to work, and we set up the guardrails that make them safe to keep: approved sources, review steps and least-privilege environments. That can be a one-off policy and setup for your dev team, or ongoing AI consulting. For agents beyond code, see how we approach business automation.
Whether you're in the Sydney CBD, Parramatta, the Northern Beaches or anywhere else in Australia, email info@ailabaustralia.com with the agents and skills your team uses, and we'll tell you honestly what we'd keep, what we'd pin and what we'd remove.
Get expert guidance on implementing ai tools solutions for your Australian business
