Ponytail

Ponytail is an open-source skill that makes Claude Code and other coding agents write the smallest complete change, for developers tired of bloated diffs.

What it changes: the agent stops at the first rung that holds

Ask an agent for a date picker and it will often install a library or build a whole calendar by hand. Ponytail puts a short ladder in front of that habit. After the agent has read the code the change touches, it works down these questions and stops at the first one that answers the task.

• Does this need to exist? If not, skip it and say so in one line.
• Is it already in this codebase? Reuse it the way the surrounding code does.
• Does the standard library or the platform do it? Use that.
• Does an installed dependency solve it? Use it, and never add one for a few lines.
• Can it be one line a reader gets at a glance? Write the one line.
• Only then: the minimum code that works, plus one small test if the logic needs it.

The project's own example is the date picker. Without the skill the agent built a calendar by hand in 335 lines. With Ponytail 5 it wrote one 10-line file that reuses the repo's existing input with the browser's built-in date type. That is the author's median of five runs, not something Seailife reproduced.

What it refuses to cut

• Validation at trust boundaries, error handling that prevents data loss, security and accessibility stay in.
• Logic with a branch, a loop, a parser, money or security leaves one small test behind.
• Every reply ends with one or two lines naming what was skipped or not checked, and any risk you should know about.
• A deliberate shortcut gets a shortcut: comment, and /ponytail-debt collects those comments into a ledger so "later" does not become "never".

Levels and commands

• /ponytail lite: builds what you asked and names the smaller option in one line.
• /ponytail (full, the default): the smallest complete change, with a check where the logic needs one.
• /ponytail ultra: also questions the request and pushes back before building. /ponytail off turns it off.
• /ponytail-review: reviews the diff for bugs, security, real load, missing tests, speed and what to cut. It reads the code your change touches, not only the diff, and each finding says what goes wrong and how to fix it.
• /ponytail-audit: the same review for the whole repo, most important first.
• /ponytail-gain shows the benchmark scoreboard and /ponytail-help prints the quick reference.

Install, and where it actually works

• Claude Code: send /plugin marketplace add DietrichGebert/ponytail, then /plugin install ponytail@ponytail as two separate prompts. The install only works when they are sent separately.
• Codex: codex plugin marketplace add DietrichGebert/ponytail and codex plugin add ponytail@ponytail, then trust its two lifecycle hooks in /hooks and start a new thread.
• Copilot, Gemini and pi have one-line installs on the site. Windsurf, Cline, Kiro, OpenCode, Zed and others are covered step by step in INSTALL.md in the repo.
• The README lists 20 agents. Slash commands need a skill-capable host; Cursor with the hooks gets level switching only, and instruction-only adapters such as Windsurf, Cline, Copilot and Kiro load the always-on rules without commands.
• The Claude Code and Codex plugins and the Cursor hooks run two small Node.js hooks, so node has to be on your PATH. Without it the skills still work but each hook call prints a harmless error.
• Install only from DietrichGebert/ponytail on GitHub or @dietrichgebert/ponytail on npm. The author says a copy that ships .exe or .dll files is not theirs.

The numbers, and what they cover

• Against the same agent with no skill, Ponytail 5 reports 53% fewer lines of code, 45% fewer output tokens, 26% lower cost and 41% less time.
• The run used Opus 5.5 in headless Claude Code on 39 tasks with five runs each: 12 feature tickets in a real FastAPI and React repo, 18 tasks with hidden checks, and 9 open "build me" or "show me" requests.
• Bash was disabled for every arm, so the benchmark measures what the agent writes, not how well it debugs.
• Hidden correctness and safety checks passed 96% with no skill, 96% with the previous Ponytail and 97% with Ponytail 5. The author also reports that 98% of risky logic ships with a test, against 68% without the skill.
• The site card says "5,700+ Claude Code sessions", while the final results file lists 585 sessions for this run. Read the headline numbers as the author's own measurement of one model in one tool, not an independent test.

Claude Code extras: the band and the board

• Since 5.3, Claude Code shows a band above the prompt with the current task's cost, tokens, agent time, how often existing code was reused and whether tests passed after the last change.
• The board is a local page with cost, agent time and tokens over time for every task across your projects. By default the data stays in ~/.config/ponytail/band/ on your machine, and /ponytail band import adds your past Claude Code sessions.
• The band is Claude Code only. Other agents get the same rules without it.

Who is behind it, and how fresh it is

• The repository belongs to a GitHub user, DietrichGebert. It was created on 2026-06-12, is MIT licensed, and had about 186,000 stars on 2026-11-07.
• The latest release, v5.3.3, is dated 2026-10-29, after v5.3.2 and v5.3.1 on 2026-10-16 and v5.2.0 on 2026-10-13.
• The README names two apps as already built with Ponytail, Retriever and PhotoReunion, and lists GreenPT as a sponsor. No bio or company is given for the author.
• The prompt itself is a single file, skills/ponytail/SKILL.md, and the compact version for agents that read a rules file is AGENTS.md, so you can read exactly what it tells your agent before installing.