vnhax
DietrichGebert / ponytail
View on GitHub ↗

Independent review by VNHAX. Not affiliated with or endorsed by the project's authors.

PONYTAIL 5-53% CODE (PROJECT BENCHMARK)CLAUDE CODE · CODEX · AGENTS.MDNEVER CUTS SECURITY OR TESTSMIT LICENSE
Ponytail: a coding agent replaces a long, over-built implementation with a short one that reuses existing code
DietrichGebert / ponytail — "He says nothing. He writes one line. It works."

Ponytail: Stop AI Coding Agents From Writing More Code Than the Task Needs

A skill that puts a strict senior developer inside your coding agent.
Before writing anything, Ponytail reads the code the change touches, reuses what exists and leans on what the platform already provides. It never cuts validation, error handling, security or accessibility.

-53%
Code written
Ponytail 5 vs same agent without it (project benchmark)
-41%
Agent time
39 tasks, Opus 5.5, 5 runs each (project benchmark)
98%
Risky logic shipped with a test
vs 68% without Ponytail (project benchmark)
HOW IT WORKS

How Ponytail Decides What to Write

From the README, the order of work for every task:

1. Understand

Read the code the change touches and trace the real data flow. Lazy about the solution, never about reading.

2. Reuse

Look for something that already exists in the repo, such as an existing Input component.

3. Use the platform

Prefer what the browser or language already provides, like a native date picker.

4. Write the minimum

Only then write new code, keeping validation, error handling, security and accessibility.

5. Report

Leave a small test for risky logic and end with what was skipped or not checked.

Commands

CommandWhat It Does
/ponytail [lite | full | ultra | off]Set the intensity or turn it off. With no argument it reports or switches on the default level.
/ponytail-reviewReviews the current diff like the senior dev who gets paged: bugs, security, real load, missing tests, slow paths, and what to cut.
/ponytail-auditThe same checks for the whole repository, most important first.
/ponytail-debtCollects the shortcut: comments you deferred into a ledger.
/ponytail-gainShows the measured impact scoreboard from the benchmark.
/ponytail bandClaude Code only: a band above the prompt with cost, tokens, agent time, reuse and test status.

The Benchmark, Read Carefully

All figures are from the Ponytail 5 README, measured by the project in Claude Code with and without the skill.

MeasureResult
Code written-53%
Agent time-41%
Cost-26%
Tokens-45%
Risky logic shipped with a test98% with Ponytail vs 68% without
Setup39 tasks including a real FastAPI + React repo, Opus 5.5, 5 runs each
Install

Install in Claude Code

/plugin marketplace add DietrichGebert/ponytail

Then, as a separate prompt, run /plugin install ponytail@ponytail. Ponytail is active in every session afterwards.

More Commands:
Codex:
codex plugin marketplace add DietrichGebert/ponytail && codex plugin add ponytail@ponytail
Then in Codex:
Open /hooks, trust the two lifecycle hooks, start a new thread
Any other agent:
Copy AGENTS.md into your project (see INSTALL.md)
Set a default level (optional):
PONYTAIL_DEFAULT_MODE=lite

Why Less Code Is Not the Same as Cutting Corners

The README is explicit that the rule was never "fewest tokens". It is to write only what the task needs, and never cut validation, error handling, security or accessibility. Lower cost and faster runs are a side effect of not building what nobody asked for.

One Prompt, Many Agents

Ponytail is a single prompt (skills/ponytail/SKILL.md) plus a compact AGENTS.md for agents that read a rules file. Everything else in the repository just loads that prompt into different agents, which is why it works across Claude Code, Codex, Copilot, Cursor, OpenCode, Gemini and others.

HONEST NOTES

Before You Install

Points worth knowing, mostly from the README:

  • Vendor benchmark: The -53% and -41% figures are the project's own measurements on its own task set. Check the effect on your codebase.
  • Official sources only: Install only from DietrichGebert/ponytail on GitHub or @dietrichgebert/ponytail on npm. It never ships .exe or .dll files.
  • Agent support varies: Commands need a skill-capable host; some features, like the band, are Claude Code only.
🧩 Works With

Claude Code and Codex natively; Copilot, Cursor, OpenCode, Gemini and others via AGENTS.md.

⚖️ License

MIT License.

📘 Docs

README and INSTALL.md in the repository, plus examples in examples/.

Frequently Asked Questions

How do I install Ponytail in Claude Code?

Run /plugin marketplace add DietrichGebert/ponytail and then /plugin install ponytail@ponytail as two separate prompts. For Codex, use the codex plugin marketplace add and codex plugin add commands, then trust its two hooks in /hooks.

Does Ponytail need a config file?

No. An optional ~/.config/ponytail/config.json or the PONYTAIL_DEFAULT_MODE environment variable can set the default level, but nothing is required.

Will Ponytail make my agent skip important checks?

It is designed not to. The README says trust-boundary validation, data-loss handling, security and accessibility are never on the chopping block, and logic with branches, loops, parsing, money or security gets a small test.

Where should I download Ponytail from?

Only from DietrichGebert/ponytail on GitHub or @dietrichgebert/ponytail on npm. The author says it never ships .exe or .dll files, so a copy that does is not genuine.