AgentSkillsDir home
Submit
Ponytail project image
Claude skills

Ponytail Skill for Coding Agents · DietrichGebert

The Ponytail skill from DietrichGebert/ponytail (MIT) puts a “lazy senior” developer inside your coding agent: solve the whole task with the smallest complete change, then say what was skipped and any risk. Pick a level, lite, full or ultra, and use /ponytail-review on a diff or /ponytail-audit on the whole repo.

Not a hairstyle tutorial, not another CLAUDE.md pep talk and not a hosted review service: AgentSkillsDir does not run Ponytail for you. The code and token savings in its README are upstream benchmarks, not reproduced here. This is a source-linked directory entry, not ponytail.dev.

  1. Prompt 1: add the marketplace

    Send it on its own.

    /plugin marketplace add DietrichGebert/ponytail
  2. Prompt 2: install the plugin

    Then open a new session; full is on by default.

    /plugin install ponytail@ponytail
  • Two prompts, one after the other
  • Both lines pasted into one prompt — upstream says the install will not work
lite/ponytail lite

Builds what you asked, then names the smaller option in one line.

full · default/ponytail full

The smallest complete change, by the rules in SKILL.md.

ultra/ponytail ultra

Also pushes back on parts of the request the need does not justify.

off/ponytail off

Back to your agent’s normal behaviour.

Codex, Cursor or another agent?

Upstream README commands: AgentSkillsDir has not run them yet.

What it is, and what it is not

Ponytail is one prompt, skills/ponytail/SKILL.md, plus adapters that load it into about twenty agents. Before writing code the agent reads the task and the code it touches, then stops at the first rung of a ladder that holds:

  1. Does it need to exist? Skip features nobody asked for, and name them.
  2. Is it already in this codebase? Reuse the helper, component or pattern.
  3. Does the standard library or the platform do it?
  4. Does an installed dependency do it? Never add one for a few lines.
  5. Can it be one line a reader gets at a glance?
  6. Only then: the minimum code that works, plus one small test if it has logic.

It is lazy about the solution, not about the change: callers, tests and fixtures the change breaks still get fixed, and validation at trust boundaries, data-loss handling, security and accessibility are never cut. Every reply ends with one or two lines on what was skipped or not checked.

It is not a hosted runtime, not a code generator of its own and not a hairstyle guide. It changes how the agent you already use decides what to write.

Install for your assistant

Commands follow the DietrichGebert/ponytail README and INSTALL.md, fetched 1 Nov 2026. AgentSkillsDir has not run them yet.

Needs Node: the Claude Code and Codex plugins and the Cursor hooks run two small Node.js lifecycle hooks, so node must be on the non-interactive shell’s PATH. Without it the skills still work, but each hook call prints a harmless “node: command not found”.

  • Claude Code

    then send each line as its own prompt
    /plugin marketplace add DietrichGebert/ponytail /plugin install ponytail@ponytail

    Upstream: “You have to send two separate prompts for the install to work.” The same pair works in the Claude Code desktop app’s Code tab, or via + → Plugins → Add plugin.

  • Codex

    then open /hooks, trust both, new thread
    codex plugin marketplace add DietrichGebert/ponytail codex plugin add ponytail@ponytail

    Run codex, open /hooks, review and trust the two lifecycle hooks, then start a new thread. The same install covers the Codex desktop app after a restart. In Codex the commands are namespaced: $ponytail:ponytail-review.

  • Cursor (hooks, needs Node)

    then new chat; type /ponytail ultra as a message
    git clone https://github.com/DietrichGebert/ponytail node ponytail/scripts/cursor-hooks.js install

    Merges two hooks into ~/.cursor/hooks.json (add --project for <project>/.cursor/hooks.json) and keeps hooks you already have. Leave the checkout where it is. Use the hooks or the .cursor/rules/ponytail.mdc rule file, not both: while the rule is present the hooks inject nothing. Subagents and cloud agents run without the ruleset.

  • skills CLI (skills only)

    then add AGENTS.md for always-on rules
    npx skills add DietrichGebert/ponytail

    Copies the six skills into your agent’s skills folder. Pick the agent with --agent, take all six with --skill '*', add --global for your user. Skills only: for the always-on ruleset also add AGENTS.md or use a plugin.

Other agents

GitHub Copilot CLI
copilot plugin marketplace add DietrichGebert/ponytail copilot plugin install ponytail@ponytail
Commands are namespaced: /ponytail:ponytail ultra.
OpenCode
opencode plugin add @dietrichgebert/ponytail
Injects the ruleset every turn and adds the levels.
Gemini CLI
gemini extensions install https://github.com/DietrichGebert/ponytail
Always-on context plus the /ponytail commands.
Windsurf, Cline, Aider, Zed, Copilot Chat …
AGENTS.md
Copy AGENTS.md or the matching rules file into your project. Always-on rules, no commands.

Full matrix (Pi, Kimi Code, Devin CLI, Grok Build, Goose, Kiro and more) and uninstall steps: INSTALL.md. Fallback for any agent: copy AGENTS.md into your project.

Upstream: only install from DietrichGebert/ponytail on GitHub or @dietrichgebert/ponytail on npm. It never ships .exe or .dll files; a copy that does is not official.

Levels and commands

Ponytail is active every session at full until you change it. Set a different default with the PONYTAIL_DEFAULT_MODE env var (lite, full, ultra or off) or defaultMode in ~/.config/ponytail/config.json. Not yet verified by AgentSkillsDir.

/ponytail [lite|full|ultra|off]
Set the level or turn it off. No argument turns it on at the default level, or reports the current one.
/ponytail-review
Reviews the current diff and the code it touches: bugs, security, real load, risky code without a test, slow paths, what to cut. Narrow it with uncommitted, staged, branch or a PR link.
/ponytail-audit
The same checks on the whole repository, ranked, with what to fix first.
/ponytail band
Claude Code only: show or hide the band above the prompt (cost, tokens, agent time, reuse, tests).
/ponytail-debt
Collects the shortcut: comments you deferred into a ledger.
/ponytail-gain · /ponytail-help
Upstream benchmark scoreboard · quick reference.

Review or audit? /ponytail-review looks at the change you are making; /ponytail-audit maps the whole repo first and ranks what to fix. Both run inside your agent, with your model and your bill.

The band (new in 5.3) sits above the prompt in Claude Code and shows task cost, tokens, agent time, code reuse and whether tests passed; its buttons switch level or run review and audit. It needs Claude Code 2.1.290 or later in the terminal or desktop app, and does not show in the VS Code chat panel, cloud sessions or claude -p. Data stays local in ~/.config/ponytail/band/.

Cursor with hooks gets level switching only, typed as a plain message. Rule-file installs (Windsurf, Cline, Copilot Chat …) load the always-on rules without any commands.

About the savings numbers

The README headline is −53% code, −41% time, −26% cost and −45% tokens, with 98% of risky logic shipping with a test (68% without). These are upstream benchmark claims: Claude Code, 39 tasks, 5 runs each. AgentSkillsDir has not reproduced them. Read the benchmark method and limits and measure on your own repo.

Ponytail and grill-me

Both are behaviour skills, but they act at different moments. grill-me interviews you before anything is decided and writes no code; Ponytail shapes the code once you know what you want. They are separate entries.

Before you install

  • SKILL.md and AGENTS.md are plain written instructions. Read them on GitHub first.
  • The plugins add lifecycle hooks that run Node on every session, and inject the ruleset into subagents (scope it with PONYTAIL_SUBAGENT_MATCHER). Review the hooks before you trust them.
  • Uninstalling the plugin leaves a little state behind (mode flag, config, statusline). Upstream ships scripts/uninstall.js; run it before removing the plugin.
  • Listing is not a security review. AgentSkillsDir has not reviewed this code.

FAQ

What is the Ponytail skill?

An open-source agent skill and plugin from DietrichGebert/ponytail (MIT) that makes coding agents work like a lazy senior developer: the smallest change that fully solves the task, then a note on what was skipped and any risk. It has lite, full and ultra levels plus /ponytail-review and /ponytail-audit. It is not a hairstyle tutorial.

How do I install Ponytail in Claude Code?

Send two separate prompts: /plugin marketplace add DietrichGebert/ponytail, then /plugin install ponytail@ponytail. Upstream says the install only works as two prompts. Fetched 1 Nov 2026; not yet verified by AgentSkillsDir.

How do I install Ponytail for Codex or Cursor?

Codex: codex plugin marketplace add DietrichGebert/ponytail, then codex plugin add ponytail@ponytail; open /hooks, trust both hooks, start a new thread. Cursor: clone the repo and run node ponytail/scripts/cursor-hooks.js install (or copy the Cursor rule file instead). Not yet verified by AgentSkillsDir.

Why do I see “node: command not found”?

The Claude Code and Codex plugins and the Cursor hooks run two small Node.js lifecycle hooks. Put node on the non-interactive shell’s PATH (Nix and nvm users take note). Per INSTALL.md the skills still work without it and the error is harmless.

lite, full or ultra?

lite builds what you asked and names a smaller option. full, the default, applies the smallest-complete-change rules. ultra also pushes back on parts of the request the need does not justify. off turns it off. Set a default with PONYTAIL_DEFAULT_MODE.

/ponytail-review or /ponytail-audit?

Review checks the current change and the code it touches. Audit runs the same checks on the whole repository and ranks what to fix first. Both run inside your own agent; AgentSkillsDir does not host either.

Are the −53% code and −45% token numbers guaranteed?

No. They are upstream benchmark claims from the README (Claude Code, 39 tasks, 5 runs each). AgentSkillsDir has not reproduced them; results depend on your model, repo and tasks.

Is Ponytail still useful on newer models?

It is debated in the community. Newer models may already write less, but Ponytail also changes the reply format, the review and audit commands and the levels. Try lite or full on your own stack and compare.

Does AgentSkillsDir run Ponytail or sell a review service?

No. This is a source-linked directory entry. Nothing installs or runs from this page, and listing is not a security review.

Are the star count and version current?

They are a dated snapshot from the GitHub API and npm; see the last-checked time on the source card.

Corrections and suggestions

Spotted something out of date, or want to suggest a skill? Email a correction. To suggest a resource, submit it for manual review.