ponytail

Ponytail is a skill, plugin, and ruleset for AI coding agents, designed to push them toward writing less and simpler code. Its tagline is the "lazy senior dev": instead of generating a wrapper component, a stylesheet, and a dependency for something like a date picker, the agent is guided to use an existing native HTML input (`<input type="date">`) when that is all the task needs. ## How it works Before writing code, the agent works through a ladder and stops at the first rung that holds: 1. Does this need to exist? If no, skip it (YAGNI). 2. Already in this codebase? Reuse it, don't rewrite. 3. Stdlib does it? Use it. 4. Native platform feature? Use it. 5. Installed dependency? Use it. 6. One line? Write one line. 7. Only then: the minimum that works. The project emphasizes that the ladder runs after the agent understands the problem, not instead of it: the agent should still read the code and trace the real flow. It also stresses that "lazy" is not "negligent": validation, error handling, security, and accessibility are never sacrificed. ## Benchmark claims The README reports an agentic benchmark using a headless Claude Code session editing a real FastAPI + React template, scored on the git diff across twelve feature tasks (Haiku 4.5, n=4). It claims ponytail reduced LOC by ~54% (up to 94% in over-build cases) while also reducing tokens, cost, and time, and remaining at 100% on a safety tier. It includes older single-shot numbers showing 80-94% less code, with a caveat that part of that gap came from a conversational baseline artifact. The project links to full writeups and reproduction instructions rather than presenting the numbers as the only evidence. ## Install and support Ponytail installs as a plugin or ruleset across a wide range of AI coding hosts, including: - Claude Code and Codex via plugin marketplace commands - GitHub Copilot CLI via plugin install - Gemini CLI and Antigravity CLI via extension install - OpenCode via a plugin entry in `opencode.json` - Pi, Hermes Agent, Devin CLI, OpenClaw, Grok Build, and Qoder via their own package mechanisms - Cursor, Windsurf, Cline, Kiro, Zed, CodeWhale, Swival, and others by copying a rules file such as `AGENTS.md` or the relevant rules directory It ships as a single repo with shared `hooks/` and `skills/` directories plus agent-specific adapters. ## Commands Skill-capable hosts expose a handful of slash commands: - `/ponytail [lite | full | ultra | off]` - set intensity or turn it off - `/ponytail-review` - review the current diff for over-engineering and return a delete-list - `/ponytail-audit` - audit the whole repo for over-engineering - `/ponytail-debt` - collect deferred "ponytail:" shortcuts into a ledger - `/ponytail-gain` - show measured impact from the benchmark - `/ponytail-help` - quick reference Levels can be controlled per session, per subagent via the `PONYTAIL_SUBAGENT_MATCHER` env var, or by default via the `PONYTAIL_DEFAULT_MODE` env var or a small config file. ## Development The repo includes scripts to keep the compact rule text in sync across all agent copies (`node scripts/check-rule-copies.js`), generate the OpenClaw skill package, publish skills to ClawHub, and run the correctness benchmark (Python is used for email and CSV checks). ## License MIT.