📊 See DietrichGebert/ponytail on AI Rank

📱 Get the AI Rank app

Ponytail Claude is a third-party plugin that loads a "lazy senior developer" ruleset into coding agents. The public project is DietrichGebert/ponytail on GitHub, MIT licensed. It is not an Anthropic product. The README presents it as one prompt, skills/ponytail/SKILL.md, plus adapters that inject that prompt into different hosts. GitHub's repository API, read on November 6, 2026, listed 185,853 stars and 10,283 forks. Those counts move; prefer the live repository page.

→ Project ranking: DietrichGebert/ponytail on AI Rank — the live page showed about 185.5k stars, 10.3k forks, and a #38 today marker when opened on November 6, 2026. Re-check the live page.

What Ponytail Claude Is

The skill file describes a coding stance: solve the whole task with the least new code, reuse what the repo already has, and end the reply with what was skipped or not checked plus any risk the user should know. It says to use the skill on coding work (writing, fixing, refactoring, reviewing, choosing dependencies) and when someone asks for the simplest path or complains about over-engineering.

The same file tells the agent not to cut validation at trust boundaries, error handling that prevents data loss, security, accessibility, hardware calibration, or anything the user asked for. New non-trivial logic (a branch, a loop, a parser, money or security, or a new script or app) is supposed to leave one small test or an assert-based self-check.

Ponytail 5 is the current README headline. The repository was created on June 12, 2026. The latest commit visible through GitHub's API when checked pointed to October 29, 2026 documentation work. The project also ships review, audit, debt, gain, and help skills beside the main mode.

Read the files before you turn it on. Anthropic's Agent Skills overview says to prefer trusted sources, or to audit bundled files if the source is unknown, because a skill can include instructions and code. Skills you upload or someone shares with you are not reviewed by Anthropic.

How to Install Ponytail Claude

The steps below are copied from the project's README and INSTALL.md. They are the author's documented commands, not a live install run for this page.

For Claude Code, the README asks for two separate prompts:

/plugin marketplace add DietrichGebert/ponytail

/plugin install ponytail@ponytail

INSTALL.md repeats that the two prompts must be sent separately. The same document says the Claude Code desktop app's Code tab accepts those /plugin commands, or you can use Plugins → Add plugin after configuring the marketplace.

For Codex, the README lists:

codex plugin marketplace add DietrichGebert/ponytail
codex plugin add ponytail@ponytail

Then open /hooks in Codex, trust the two lifecycle hooks, and start a new thread.

For any other agent, the README says to copy AGENTS.md into the project, or ask the agent to install skills/ponytail/SKILL.md as a skill. Host-by-host commands for Copilot, Cursor, OpenCode, Gemini, and others live in INSTALL.md. This page does not replay that full list.

INSTALL.md notes that the Claude Code and Codex plugins (and the Cursor hooks) run two small Node.js lifecycle hooks, so node needs to be on the PATH. If it is missing, the document says the skills still work, but each hook call shows a harmless node: command not found error.

The README's trust line is strict: only install ponytail from DietrichGebert/ponytail on GitHub or @dietrichgebert/ponytail on npm. It says the project never ships .exe or .dll files, and a copy that does is not the author's.

Default intensity is full. INSTALL.md says you can change that with the PONYTAIL_DEFAULT_MODE environment variable (lite, full, ultra, or off) or a defaultMode field in ~/.config/ponytail/config.json (Windows: %APPDATA%\ponytail\config.json). Resolution is environment variable, then the config file, then full.

To remove it from Claude Code, INSTALL.md lists /plugin remove ponytail.

Lite Full Ultra Levels

Lite full ultra are the three intensities in skills/ponytail/SKILL.md and the help skill. They are real switches in the repo, not a rumor.

Level What the skill file says
lite Build what was asked. Name the smaller option in one line and let the user pick.
full The main rules. Default.
ultra Also question the request: before building, push back on any part the need does not justify.

The help skill adds that /ponytail lite, /ponytail (full), and /ponytail ultra set the level, and that the choice sticks until you change it or the session ends. /ponytail off, or saying "stop ponytail" or "normal mode", turns it off.

/ponytail with no argument switches it on at the default level if it is off, and otherwise reports the current level.

README also documents Claude Code-only band commands: /ponytail band hides or shows a status band above the prompt, and /ponytail band import adds past Claude Code sessions to a local board. The README says the band needs Claude Code 2.1.290 or later, in the terminal or the desktop app, and that it does not show in the VS Code chat panel, cloud sessions, or claude -p. By default, band data stays on the machine in ~/.config/ponytail/band/.

What the Project's Benchmark Claims

The README headline for Ponytail 5 is: -53% code, -41% time, -26% cost, -45% tokens, and 98% of "risky logic" shipping with a test versus 68% without Ponytail. Those figures are the project's own claims. They come from benchmarks/results/2026-10-07-agentic.md.

That write-up describes the setup as headless Claude Code sessions (claude -p) with Opus 5.5 and Claude Code CLI 2.1.288, 39 tasks, and 5 runs each. The tasks include 12 feature tickets in a FastAPI + React repo (full-stack-fastapi-template), 18 surgical tasks with hidden checks, and 9 smaller "build me an app" or "show me an example" prompts. Bash was disabled for every arm. The author compares Ponytail 5, Ponytail v4.13.0, and the same agent with no skill.

Against no skill, the file reports geometric-mean ratios (median of five runs per task) of -53% lines of code (95% bootstrap interval -63 to -41), -41% gzip of delivered code, -45% output tokens, -26% cost (-36 to -16), and -41% time. It also reports cheaper results on 30 of 38 tasks, hidden checks passed on 87 of 90 surgical runs for Ponytail 5 (versus 86 of 90 with no skill), and "a test where the logic needs one" at 98% versus 68% with no skill. On six named security tasks, the file says all three arms passed 30 of 30 runs.

Treat those numbers as the author's benchmark, scored in that harness, on that date. They are not an independent re-run, and they are not a promise that your repo will see the same cut in code, time, or cost.

The README's date-picker example makes the same point in a smaller story: without Ponytail the agent built or installed a large date picker (335 lines in that example); Ponytail 5 reused an existing Input component and the browser's native date picker (10 lines). That example also comes from the project, not from a separate test.

Commands and Where Ponytail Applies

The README command table lists:

Command README role
/ponytail [lite | full | ultra | off] Set intensity, or turn it off
/ponytail band [on | off] Claude Code only: hide or show the band
/ponytail band import Claude Code only: import past session costs into the local board
/ponytail-review Review the current diff for bugs, security, load, missing tests, speed, and what to cut
/ponytail-audit The same review for the whole repo, ranked
/ponytail-debt Collect shortcut: comments into a ledger
/ponytail-gain Show the benchmark impact scoreboard
/ponytail-help Quick reference

The README says those commands need a skill-capable host, and names Claude Code, Codex, Devin CLI, OpenCode, Gemini, Antigravity CLI, pi, Hermes Agent, Qoder, and Grok Build. In Codex CLI it says to invoke skills as $ponytail:ponytail-review. Cursor with the hooks gets /ponytail level switching typed as a plain message. Instruction-only adapters (a Cursor rule file, Windsurf, Cline, Copilot, Kiro, or an AGENTS.md fallback) load the rules without the command set.

The skill is aimed at coding work. It is not documented as a general-purpose chatbot persona for non-code tasks.

A related community-skill explainer on AI Rank is Instagram Agent Skill. That pack writes social captions; it is not Ponytail.

FAQ

Is Ponytail Claude an official Anthropic tool?

No. It is a third-party GitHub project by DietrichGebert. Anthropic publishes the skill format; this repository fills that format with its own rules.

How do I install Ponytail for Claude Code?

Send two separate prompts: /plugin marketplace add DietrichGebert/ponytail, then /plugin install ponytail@ponytail. That sequence is from the README. Read the files first, and only install from the author's GitHub repo or the named npm package.

What is the difference between lite, full, and ultra?

Lite builds what you asked and names a smaller option. Full applies the main "smallest complete change" rules and is the default. Ultra also pushes back on parts of the request before building. Those definitions come from the skill files.

Do the benchmark percentages apply to my project?

The author measured them in a specific Claude Code harness on October 7, 2026. Your editor, model, and codebase can differ. Read the method file if you want the limits.

Does Ponytail need a config file?

The README FAQ says no. A config file or PONYTAIL_DEFAULT_MODE can change the default level, but nothing is required.

Sources

Not affiliated with DietrichGebert or Anthropic. Repository contents change; read the files you install.

Keep exploring: see which open-source AI projects are rising today in the open-source rankings on AI Rank.

Or 📱 get the AI Rank app to follow them on your phone.

Prepared with automated editorial assistance and checked against the repository files and dated AI Rank data on November 6, 2026. We did not run our own benchmarks for this article.