Refactoring · Blog Post

Ponytail: the "lazy senior developer" ruleset for coding agents

Ask a coding agent for a date picker and it may install a library or hand-build a calendar. Ponytail is one prompt that tells the agent to stop first and check what already exists: the codebase, the standard library, the platform. It also ships a review, a repo audit, and a ledger for the shortcuts it leaves behind.

By Mehadi Hasan 8 min read

What it is

Ponytail is an open-source skill pack by Dietrich Gebert. The repo describes it as making your agent "think like the laziest senior dev in the room." The idea is the long-time colleague who reads your fifty lines and replaces them with one.

Under the jokes, it is a set of written rules. The plugin manifest sums it up plainly: the simplest, shortest solution that actually works, YAGNI, standard library first, no abstractions nobody asked for. The README is clear that the goal is not the fewest tokens. The rule is to write only what the task needs, and never cut validation, error handling, security or accessibility.

The whole thing is one prompt, skills/ponytail/SKILL.md, plus a compact AGENTS.md for agents that read a rules file. Everything else in the repo exists to load that prompt into different agents. The repo was created on June 12, 2026, the current version is 5.3.3, and it is MIT licensed.

What's inside

192k+ GitHub stars (Nov 17, 2026)
6 Skills in the pack
3 Levels: lite, full, ultra
20 Agents listed as supported

The star count is 192,809 as of November 17, 2026. The "20 agents" figure is the README's own badge. Six skills ship in the skills/ folder:

Skill What it does
ponytail The always-on ruleset. Smallest complete change, one small test where logic needs it, and a closing line on what was skipped and any risk.
ponytail-review Reviews a diff, staged changes, a branch or a PR link. Reads the code the change touches, not only the diff. Changes nothing.
ponytail-audit The same review for a whole repo or a named folder. Ranked into must fix, should fix, nice to have. At most 20 findings. Changes nothing.
ponytail-debt Greps every shortcut: comment (older code may say ponytail:) into a ledger with the limit and the upgrade trigger. Changes nothing.
ponytail-gain Prints the published benchmark scoreboard. It is told never to invent a per-repo savings number.
ponytail-help A reference card for levels, commands, settings and updating.

In Claude Code there is also "the band", added in 5.3. It sits above the prompt and shows what the current task cost, tokens, agent time, how often existing code was reused, and whether tests passed after the last change. A local board page collects every task across projects. By default that data stays on your machine in ~/.config/ponytail/band/. The README says the band needs Claude Code 2.1.290 or later and does not show in the VS Code chat panel, cloud sessions or claude -p.

How it works: the ladder

The core skill starts with reading. Before writing anything, the agent must read the task and the code it touches, and list every place the change must reach: callers, tests, fixtures, config, exports. Then it takes the first option that fully works:

  1. Does it need to exist? Skip features and options nobody asked for, and name them in one line.
  2. Is it already in this codebase? Use the existing helper, component or pattern.
  3. Does the standard library or the platform do it? Use that, unless the project has its own.
  4. Is there an installed dependency for it? Use it. Never add a dependency for a few lines.
  5. Can it be one line a reader gets at a glance? One line.
  6. Otherwise, the minimum code that works.

The README's own phrase for this is "lazy about the solution, never about reading." A few rules stop "lazy" from turning into "careless." A bug fix must grep every caller and fix the root cause once in the shared code. Code that gets moved keeps its error handling. New logic with a branch, a loop, a parser, money or security must leave one small test or an assert-based self-check behind. And the skill lists what it may never cut: validation at trust boundaries, error handling that prevents data loss, security, accessibility, and anything the user asked for.

When the agent takes a shortcut with a known limit, it writes a comment in a fixed form. That form is what /ponytail-debt reads later:

Shortcut comment convention

shortcut: <the limit>, <when to upgrade>

Every reply ends with one or two lines on what was skipped or not checked, and any risk you should know. Three levels change how hard it pushes. Lite builds what you asked and names the smaller option in one line. Full is the default and applies the rules above. Ultra also questions the request and pushes back on any part the need does not justify.

Installation

In Claude Code, send these as two separate prompts. INSTALL.md says the install does not work if you combine them:

Claude Code: plugin marketplace

/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail

For Codex, run the two commands below, then open /hooks in Codex, trust its two lifecycle hooks, and start a new thread:

Codex

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

For Gemini CLI there is an extension install:

Gemini CLI

gemini extensions install https://github.com/DietrichGebert/ponytail

INSTALL.md also has commands for GitHub Copilot CLI, OpenCode, Devin CLI, Grok Build, Hermes Agent, Kimi Code, Goose, Antigravity CLI, OpenClaw and others. Cursor, Windsurf, Cline, Copilot Chat in the editor, Aider and Kiro get the ruleset by copying a rules file from the repo. The skills CLI route, npx skills add DietrichGebert/ponytail, copies the six skills only. For the always-on rules you still need AGENTS.md or a plugin.

Two practical notes from INSTALL.md. The Claude Code and Codex plugins run two small Node.js hooks, so node must be on your PATH, or you get a harmless "command not found" message on every hook call. And the uninstall commands leave a few state files behind. The repo ships scripts/uninstall.js for those, and says to run it before removing the plugin. The README also warns to install only from DietrichGebert/ponytail on GitHub or @dietrichgebert/ponytail on npm.

Using it day to day

Once installed, it is on for every session at the full level. You just ask for work as usual. Change the level with /ponytail lite, /ponytail ultra or /ponytail off, or say "stop ponytail". To set a different default, use the PONYTAIL_DEFAULT_MODE env var or a defaultMode field in ~/.config/ponytail/config.json. No config is required. While active, the ruleset is also injected into subagents. PONYTAIL_SUBAGENT_MATCHER lets you limit that to certain agent types.

The commands you will reach for:

  • /ponytail-review before merging. Name a target in plain words: uncommitted, staged, branch, or a PR link. Each finding has four parts: what the code is, the problem, the fix, and what happens if you skip it. It ends with a verdict such as "Ship." or "fix 1 and 3 first."
  • /ponytail-audit when you inherit a repo. It maps entry points and data flow first, says which load it assumed, and ranks findings. Lean findings (delete, reuse, stdlib, yagni, merge, split) come last in priority, after bugs, security, scale, missing tests and speed.
  • /ponytail-debt now and then, so deferred shortcuts don't rot. Comments with no upgrade trigger get a no-trigger tag. You can pass your own marker word, such as /ponytail-debt TODO.

Review, audit, debt, gain and help are all report-only. Each SKILL.md says it changes no code. In Codex the commands are namespaced, for example $ponytail:ponytail-review. Rules-file-only setups (Cursor's rule, Windsurf, Cline, Copilot Chat, Kiro) get the always-on behaviour but not the commands.

The benchmark claims, read carefully

The README leads with Ponytail 5's numbers against the same agent without the skill: 53% less code, 41% less time, 26% lower cost, 45% fewer tokens. It also says 98% of risky logic shipped with a test, against 68% without. The setup was headless Claude Code on Opus 5.5, 39 tasks including a real FastAPI and React repo, five runs each. The author ran it, and the method is published in benchmarks/results/2026-10-07-agentic.md.

That results file is worth reading, because it is honest about the limits. Only 18 of the 39 tasks have hidden checks that run the code. On those, Ponytail 5 passed 87 of 90 runs and no skill passed 86 of 90. The file's own conclusion is that less code did not cost correctness or safety, but did not add any either. In a blind judge comparison of replies against no skill, the difference was not significant, with a slight lean toward the longer no-skill replies. So the defensible claim is smaller output at similar correctness, on one model and one harness.

Pricing

Free. The repo, the npm package and every skill are MIT licensed. You pay only for your agent's usage, and the point of the skill is to reduce that. The band's optional webhook can send each task's row as JSON to an address you choose, which the manifest suggests for a team board, a sheet script or n8n. It is empty by default, so nothing leaves your machine unless you set it. The README also links a waitlist for something unannounced. Nothing in the repo says what it is.

My take

The useful part is not "write less code." It is the order of checks. Most agent bloat comes from skipping rungs two and three of the ladder: not looking for the helper the repo already has, or not knowing the platform already does the job. A written rule that forces that look, plus a rule that bug fixes go to the shared function rather than one caller, targets the mistakes that show up most often in agent diffs.

The report-only skills are the other strong piece. A review and an audit that rank by bug, security, load and missing tests before style, and that need a concrete failing case for every finding, is a sensible bar. The debt ledger is a small idea done well: if the agent is going to take shortcuts, at least they are written down with a trigger to revisit.

The honest caveats. The numbers are the author's own, on one model, and the correctness gain against no skill is flat. A team with strong review habits may get less from the review skills. And "minimal" is a judgment call: on a codebase with deliberate abstractions, ultra mode's pushback may be noise. Lite exists for that.

It suits solo developers and small teams whose agents keep adding dependencies and wrappers, and anyone who wants a cost and token readout per task in Claude Code. It is a light install to try: two prompts in Claude Code, and /ponytail off if it gets in the way.

Frequently asked questions

Does "lazy" mean the agent skips validation or tests?

No. The core SKILL.md lists what it may never cut: validation at trust boundaries, error handling that prevents data loss, security, accessibility, and anything you asked for. New logic with a branch, loop, parser, money or security must leave one small test or an assert-based self-check behind.

What is the debt ledger?

When the agent takes a shortcut with a known limit, it writes a comment in the form "shortcut: <the limit>, <when to upgrade>" (older versions used "ponytail:"). /ponytail-debt greps those comments into one list with the limit and the upgrade trigger, and tags any without a trigger as no-trigger. It only reports and changes nothing.

Do the review and audit commands change my code?

No. /ponytail-review, /ponytail-audit, /ponytail-debt, /ponytail-gain and /ponytail-help are all report-only. Review and audit rank findings as must fix, should fix and nice to have, and every finding needs a concrete case where the code goes wrong.

How do I turn it down or off?

Use /ponytail lite to have it build what you asked and only suggest the smaller option, /ponytail off (or say "stop ponytail") to disable it, and /ponytail to turn it back on. Set a default with the PONYTAIL_DEFAULT_MODE env var or a defaultMode field in ~/.config/ponytail/config.json.