The hottest AI trend on GitHub tells the robot to write less code. Its own benchmark got corrected.

Ask an AI coding agent to add a button, and you might get a button, three helper classes, a config file, a retry framework, and a small essay explaining all of it. You wanted a button.

This summer, one of the fastest-growing projects on GitHub fixed that with a trick so simple it feels like cheating: it tells the robot to be lazy.

The laziest senior developer in the room

The project is called Ponytail. It was published on GitHub on 12 June 2026 by a developer who goes by DietrichGebert, and it is free and MIT licensed. InfoQ reported it passed 82,000 stars by August, and it was still adding more than a thousand stars in a day on GitHub’s trending list this week.

It is not really software. It is a skill: a bundle of written instructions that your coding agent reads before it starts working. The instructions push the agent to behave like a senior engineer who really, really does not want to write extra code. Before writing anything, the agent has to climb a ladder:

1. Does this need to exist at all?

2. Is it already somewhere in this codebase?

3. Does the standard library do it?

4. Does the platform already have a built-in feature for it?

5. Is there a dependency already installed that handles it?

6. Can it be done in one line?

7. Only then: write the smallest thing that works.

It works with more than a dozen agents, including Claude Code, Codex, Cursor, GitHub Copilot and Gemini CLI. In Claude Code, installing it takes two commands. There are also slash commands to turn the intensity up or down, to review a diff for over-engineering, and to scan a whole repo.

So does it actually work? (The plot twist)

Here is where the story gets good, because Ponytail’s first scoreboard was wrong.

The launch benchmark claimed it cut code by 80 to 94 percent. Colin Eberhardt, CTO at the software firm Scott Logic, took a look under the hood. As InfoQ tells it, the repository held thousands of lines, but only around 100 of them were real instructions, and those mostly restated the old “You Aren’t Gonna Need It” rule that programmers have preached for decades. Even better, he found that swapping Ponytail out for a single sentence asking the agent to follow that principle did better. The comparison had been unfair: the “before” agent was padding its answers with chatter, which made the “after” look magical.

The maintainer’s response is the part worth applauding. He agreed the baseline was flawed and rebuilt the test against fairer opponents: twelve real feature tasks, done by Claude Code inside real FastAPI and React projects. The honest numbers are more modest:

The big 94% figure now only applies to cases where an agent was badly over-building. Where the code was already lean, savings were close to zero. The write-ups also warn that on very wordy reasoning models, the token savings can even go the wrong way.

Still impressive. Still not magic.

One quick word from today’s sponsor

AI can build faster. Can your team decide better?

AI can draft the PRD and prototype the idea. Jira Product Discovery helps teams decide whether it belongs on the roadmap. Bring feedback and ideas together, prioritize as a team, and keep your roadmap connected to delivery in Jira.

Get Started for Free

What can go wrong

The fun part of Ponytail is that it is just a text file. The scary part of Ponytail is that it is just a text file.

• A skill is an instruction your agent obeys. Anything you install into an agent’s skills gets read as if you had typed it yourself. A good one saves you money. A malicious one could tell the agent to do something you would never approve. Treat a skill like a script you are about to run, because functionally it is one.

• Benchmarks are marketing until someone checks them. The original 94% number was the headline until somebody checked it. Ask who ran the test, against what, and whether you can repeat it.

• Less code is not always better code. Deleting a “pointless” check can delete the check that was stopping bad data from getting in. Ponytail says it protects validation, error handling, security and accessibility, but that is a promise from a rule file, not a guarantee from a compiler.

• Lazy agents still hallucinate. An agent hunting for something that “already exists” may confidently name a library that does not. Researchers at the Cloud Security Alliance’s AI Safety Initiative reported in April that about 19.7% of the 2.23 million code samples they analysed across 16 models named packages that do not exist, and 43% of those made-up names came back every time the prompt was rerun. That repeatability is what lets attackers register the fake names and wait. The trick has a name: slopsquatting.

How to enjoy the shortcut without getting burned

1. Read the skill before you install it. Open the file. It is plain text. If you would not paste it into your terminal, do not hand it to your agent.

2. Pin a version. Install a specific release and read the changes before you update, the same way you would for any dependency.

3. Verify every package your agent names. Check that it exists, how old it is, who publishes it and how many people use it before anything gets installed.

4. Run your own mini benchmark. Pick three tasks you really do, run them with and without the skill, and compare. Your repo is the only one that matters.

5. Keep a human on the diff. Shorter diffs are easier to review. That is the real win here, so use it.

The takeaway

The most interesting thing about Ponytail is not that it makes agents write less. It is that a hugely popular project had its headline number checked in public, corrected, and re-published. In a corner of tech that runs on hype, that is rare and worth copying.

If you only remember one thing: trust the tool that shows its working, not the one with the biggest number.

Facts here come from the Ponytail GitHub repository, InfoQ, Pasquale Pillitteri’s write-up, and Startup Fortune’s report on the Cloud Security Alliance research. It is a fast-moving project and details may change.