Ponytail

Rules for a coding agent, installed as a skill or plugin, that change what the agent writes: before any code it stops at the first of seven rungs that holds — it does not need to exist, it is already in the codebase, the standard library does it, a native platform feature does it, an installed dependency does it, it is one line — and only then writes the minimum that works. It ships as six skills and as plugins for the hosts that have a plugin system, with a per-session level of lite, full, ultra or off, default full, also reaching subagents. Installing takes one or two commands: /plugin marketplace add DietrichGebert/ponytail then /plugin install ponytail@ponytail in Claude Code; codex plugin marketplace add DietrichGebert/ponytail then codex plugin add ponytail@ponytail in Codex; node ponytail/scripts/cursor-hooks.js install for Cursor; or one line in opencode.json. Hosts without a plugin system copy one rules file. The default level can be pinned with PONYTAIL_DEFAULT_MODE or ~/.config/ponytail/config.json.

Who built itOf the repository’s 224 commits, 114 carry his own GitHub account and 8 more use a second email of his from another machine; the rest come from about sixty contributors, one of whom has 23 commits under the name Emeriko.

1. 01

   Seventeen days of releases, then thirty-nine days of silence

   The repository was created on 2026-06-12; its first commit is Initial commit at 00:52:37 and the last is chore: release v4.10.0 (#870) at 2026-09-14T14:34:42Z. That is 224 commits in 94 days, piled into the first two months: 155 in June, 51 in July, 4 in August, 14 in September. The 16 release titles carry the story. v1.0.0 — He ships. went out on 2026-06-12T02:43:54Z, under two hours after the first commit; v4.0.0: production grade, still lazy followed the same day, then v4.1.0: three more agents and v4.2.0: lazy in OpenCode now on 2026-06-13, and on 2026-06-15 v4.3.0: more agents, still lazy, v4.4.0: field-tested, still lazy, v4.5.0: lazy in Copilot and, four hours later, v4.6.0: help, reluctantly. v4.7.0: lazy in OpenClaw now arrived on 2026-06-16; the four releases of 2026-06-23 and 2026-06-24 are v4.8.0: comprehension first, now with an MCP server, v4.8.1: consistent versioning, v4.8.2: now on npm and v4.8.3: lazy in subagents too; and v4.8.4: lazy in Hermes now closed the run on 2026-06-29 — 14 of the 16 releases inside seventeen days. Thirty-nine days then pass before v4.9.0: 53 commits of doing less on 2026-08-07 and v4.10.0 on 2026-09-14. Around it: 148,962 stars, 8,009 forks, 356 watchers, 319 open issues against those 224 commits, about one commit per 665 stars, in a 2,666 KB JavaScript repository.

2. 02

   The product is one document and six commands

   What gets installed is small enough to quote. AGENTS.md (2,593 bytes) opens: “You are a lazy senior developer. Lazy means efficient, not careless. The best code is the code never written.” Then seven rungs, each a question: does this need to exist at all (YAGNI); is it already in this codebase; does the standard library do it; does a native platform feature cover it; does an installed dependency solve it; can it be one line; only then, the minimum that works. The ladder runs “after you understand the problem, not instead of it”. The same file names what is not on the chopping block — input validation at trust boundaries, error handling that prevents data loss, security, accessibility — and requires that non-trivial logic “leaves ONE runnable check behind”, while “trivial one-liners need no test”. Around it sit six skills. /ponytail sets the intensity to lite, full, ultra or off; /ponytail-review reads the current diff for over-engineering and returns a delete-list; /ponytail-audit does the same for a whole repository; /ponytail-debt harvests the markers left on deliberate shortcuts into a ledger, “so ‘later’ doesn’t become ‘never’”; /ponytail-gain prints the measured scoreboard; /ponytail-help lists the rest. The default is full, set by PONYTAIL_DEFAULT_MODE or defaultMode in ~/.config/ponytail/config.json.

3. 03

   Half the commits name a model as co-author

   Of the 224 commits, 112 carry a Co-authored-by trailer, and the biggest groups name a model rather than a person: Claude Opus 4.8 on 32 commits, Claude Opus 4.8 (1M context) on 31, Claude Fable 5 on 7, Claude Opus 5 on 3, and a bare Claude on 1. Four name a tool or a bot instead — Cursor, Copilot, Devin and google-labs-jules[bot], one commit each. People appear in the same trailers: Emeriko on 7, Dietrich Gebert on 5, Admin on 2, then a long tail of ones. The account breakdown carries the rest: 215 of the 224 commits have a linked account, and 114 of them are the owner’s own. By email, 103 commits come from dietrich.gebert@gmail.com and 8 from dgebert@Dietrichs-MacBook-Pro.local, a second address of his. By author name there are 70 distinct entries, among them one — Emeriko, with 23 commits — that has no matching row in the by-account table. The contributor list holds 50 names, led by Lakshya77089 with 10, ousamabenyounes with 8, then dhedhialy, hamza-ali-shahjahan and salaamdev with 4 each. What makes this more than trivia is the last line of AGENTS.md, which turns the ruleset on the repository itself: “(Yes, this file also applies to agents working on the ponytail repo itself. Especially to them.)”

4. 04

   Twenty hosts, one ruleset, eight copies of it

   The README gives install steps for Claude Code, Codex, GitHub Copilot CLI, the Pi agent harness, OpenCode, Gemini CLI, Qoder, Antigravity CLI, Hermes Agent, CodeWhale, Swival, Devin CLI, OpenClaw, Grok Build and Cursor, and names instruction-only hosts as well: Windsurf, Cline, Copilot Chat, Aider, Kiro, Zed. A badge at the top of the README says “works with 20 agents”. The tree shows what that surface costs: the same rule text is checked in eight times, once per host format — .agents/rules/ponytail.md, .clinerules/ponytail.md, .cursor/rules/ponytail.mdc (2,620 bytes), .github/copilot-instructions.md, .kiro/steering/ponytail.md (2,560), .qoder/rules/ponytail.md, .windsurf/rules/ponytail.md and AGENTS.md, most of them 2,495 bytes apiece — with scripts/check-rule-copies.js (2,981 bytes) and node scripts/check-versions.js to keep the copies and the version strings aligned. Hosts with a plugin system get code instead: hooks/ holds twelve files and 34 KB, among them ponytail-activate.js (4,601), ponytail-config.js (5,881), ponytail-instructions.js (5,487), ponytail-mode-tracker.js (6,194) and ponytail-subagent.js (2,683), plus one hook manifest per host. The same six commands exist twice, as six .opencode/command/*.md files (510 to 1,141 bytes) and six commands/*.toml files (515 to 1,152 bytes).

5. 05

   The OpenCode 2 port, settled in public

   The busiest thread in the report is not about writing less code; it is about one host changing its plugin API. Issue #941, opened 2026-09-26, reports that @dietrichgebert/ponytail@4.10.0 cannot be loaded by OpenCode v2 at all: “Plugin must export a default definition with an id and an effect or setup function. (cause: SchemaError(Expected object at ["default"]))”. The same failure was reproduced on v2.0.18 on Windows 11 with Node v24.19.0, on Linux x64 with Node v22.23.2, and on macOS arm64; #930 had already been opened against v2.0.15 and closed by its own reporter as “filed prematurely”. What follows is a queue of competing fixes: #928, closed as a duplicate of #915 and #907, then #933, #940 and #943. Reviewing them side by side in #943, a commenter lists #907, #940, #933, #864, #729 and #734 as the alternatives and calls #943 the cleanest; the same thread reports that on 2.0.20 the dual { id, setup, server } export registers 0 commands and 0 skills, while deleting the server key from the same file registers six — the log lines read commands: 2, ponytail: 0 and commands: 8, ponytail: 6. One pull request takes the other direction on size: #939 drops assets/ and the pi-extension tests from the npm package because nothing that runs from it references them, taking the published package from 997 kB to 49 kB and from 54 files to 38.

6. 06

   The benchmark and its own correction

   The README claims “~54% less code (up to 94%) · ~20% cheaper · ~27% faster · 100% safe”, measured by a headless Claude Code session editing tiangolo’s full-stack-fastapi-template — a real FastAPI + React repository — and scored on the git diff it leaves: twelve feature tickets, the same agent with and without the skill, n=4, Haiku 4.5. The table gives ponytail −54% lines of code, −22% tokens, −20% cost and −27% time against a no-skill baseline, at 100% on a separate adversarial safety tier; the terse-prose control, caveman, shows −20% / +7% / +3% / +2% and 100% safe, and a “YAGNI + one-liners” prompt shows −33% / −14% / −21% / −30% and 95% safe. The largest cuts land where an agent over-builds: a date picker going from 404 lines to 23, a color picker from 287 to 23, reaching for a native <input> instead of a component. The README also corrects its own earlier numbers: the older single-shot run, which reported “80-94% less code”, is kept in a collapsed section with the note that issue #126 “fairly pointed out that the bare-model baseline pads its answer with prose and options, so that gap is partly a conversational-baseline artifact”, and the badge says the flat figure is “the per-task ceiling, not the average”. Nine result files under benchmarks/results/ carry the dates, from 2026-06-12 to 2026-06-22, the largest a 12,226-byte agentic write-up.