claude/for
Search
Subscribe
Claude CodeUpdated

Repo radar: ponytail: lazy-senior-dev plugin for Claude Code

ponytail is a community, non-Anthropic Claude Code plugin: one write-the-least-code prompt plus review commands. MIT. Benchmarks are README claims.

ponytail (DietrichGebert/ponytail) is a community, non-Anthropic plugin for Claude Code. Per its README, it's "one prompt" (skills/ponytail/SKILL.md) that pushes the agent toward the least code that works: skip what isn't needed, reuse the codebase, then stdlib, native platform features, and existing dependencies before writing anything new. Review and audit commands come with it. The rest of the repo adapts that prompt to other agents. License MIT. Plugin manifest version 4.12.0. Checked 2026-10-05.

TL;DR. Worth a look if your agent over-builds (extra deps, wrappers, abstractions) and you want an always-on minimalism rule plus a diff-review command. Install from the official repo only. The benchmark numbers are the author's README claims, from a small run. 155,153★ is popularity, not a quality signal.

When to use / when not

Use when you already work in Claude Code, you keep deleting code the agent didn't need to write, and you'd rather install one rule than repeat "keep it simple" every session.

Skip when you want official, Anthropic-maintained building blocks. Start with the Anthropic skills marketplace and Claude skills. If you only want a minimalism rule, you can write your own skill in a few lines. Also skip it if you can't afford extra context on every turn and in every subagent (see Limits).

Install: one path (Claude Code)

The README's two commands, run as separate prompts:

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

INSTALL.md notes the plugin runs two Node.js lifecycle hooks, so node must be on your PATH, including the non-interactive shell's PATH for nvm/Nix users. Without it, the skills still work, but each hook call prints node: command not found.

Official channels (README): install only from DietrichGebert/ponytail on GitHub or @dietrichgebert/ponytail on npm. The author says it "never ships .exe or .dll files." Before you trust any plugin long term, run it through a plugin eval.

Uninstall: INSTALL.md says ponytail leaves state outside the plugin folder: ~/.claude/.ponytail-active, ~/.config/ponytail/config.json, and an optional statusLine entry in ~/.claude/settings.json. Run node scripts/uninstall.js (from the plugin files or a separate clone) first, then remove the plugin. The repo's table says /plugin remove ponytail. Claude Code's own plugin docs use /plugin uninstall, so use that form.

What's inside: README claims

The ladder. Before writing code, the agent stops at the first rung that holds: needed at all (YAGNI) → already in the codebase → stdlib → native platform feature → installed dependency → one line → only then the minimum. The README says this runs after reading the code the change touches. It also says validation, data-loss handling, security, and accessibility are never cut.

Commands (README):

CommandPurpose
/ponytail [lite|full|ultra|off]Set intensity (default full) or turn off
/ponytail-reviewReview the current diff for over-engineering
/ponytail-auditSame, whole repo
/ponytail-debtCollect deferred ponytail: shortcuts into a ledger
/ponytail-gainShow the benchmark scoreboard
/ponytail-helpQuick reference

Config: optional PONYTAIL_DEFAULT_MODE env var or ~/.config/ponytail/config.json. By default the ruleset is injected into every subagent spawned via the Agent tool. PONYTAIL_SUBAGENT_MATCHER (a regex on agent type) narrows that.

Other agents: adapter folders for Codex, Cursor, Windsurf, Cline, OpenCode, Gemini, and more. Commands only work on skill-capable hosts. Rules-file adapters get the always-on ruleset without commands.

Benchmark (author's claim, not ours): the README reports ~54% less code, ~20% lower cost, and ~27% faster vs a no-skill baseline. That's from headless Claude Code sessions on one repo (fastapi full-stack template), 12 tasks, Haiku 4.5, n=4. The author also says older 80–94% single-shot numbers were partly a baseline artifact (issue #126), and that on a terse reasoning model (GPT-5.5) cost can go the other way. We haven't reproduced any of this [GAP].

Limits

  • Not Anthropic. Listing it here isn't an endorsement.
  • A prompt, not a guarantee. It changes what the agent tends to write. You still review diffs for correctness and security.
  • Context cost. Always-on, every turn, every subagent unless you scope it.
  • Writes outside its folder. Clean up with the uninstall script before removing the plugin.
  • Stars ≠ quality. GitHub API showed 155,153★ at 08:03 CEST on 2026-10-05. That's popularity.
  • Not inspected: hook script internals and the ponytail-mcp/ folder [GAP].

Similar community-harness fiche: Repo radar: ECC.

Sources (checked 2026-10-05)

One Claude move in your inbox, every Sunday

Four minutes, tested on a real job, then back to your weekend. Free.