Ponytail
Make new code the last resort
A plugin that makes coding agents write the least code that works. Before adding anything it walks a seven-rung ladder: skip it (YAGNI), reuse what is already in the repo, use the stdlib, use a native platform API, use an already-installed dependency, try one line, and only then write the minimum. Ships for Claude Code, Codex, Copilot CLI, Gemini CLI and other harnesses. MIT.
What is Ponytail?
Ponytail is a plugin that makes AI coding agents write the least code that works. Before adding anything, the agent walks a fixed ladder: skip speculative work, reuse what is already in the repo, use the standard library or a native platform API, use an already-installed dependency, try one line, and only then write the minimum. It is for developers whose agents over-build.
Key features
- Seven-rung decision ladder enforced after the agent has read the touched code: YAGNI, reuse, stdlib, native platform feature, installed dependency, one line, then minimum
- Three intensities: lite (names the lazier alternative and lets you pick), full (ladder enforced), ultra (YAGNI extremist)
- Installs as a plugin or extension for Claude Code, Codex, Copilot CLI, Gemini CLI, the pi harness, and other listed agents (site: 14+; README badge: 20 agents)
- Chat commands to set intensity, audit the current diff for over-engineering, scan the repo for bloat, collect deferred shortcuts, and show the benchmark scoreboard
- Safety rule published on the site: validation, error handling, security and accessibility are never simplified away
- MIT licence, source at github.com/DietrichGebert/ponytail, product site ponytail.dev
- First-party agentic benchmark on a real FastAPI + React repo (12 feature tasks, Haiku 4.5): claims ~54% less code, ~22% fewer tokens, ~20% lower cost, ~27% faster vs the same agent with no skill
Who it's for
- A Claude Code or Codex user whose agent keeps installing a date-picker library onto a form the browser already handles with input type=date
- A maintainer who wants new agent output to reuse helpers already in the repo instead of rewriting them
- A team that wants a reversible intensity knob (lite / full / ultra) rather than a one-shot 'write less code' prompt
- Someone auditing an agent diff for over-engineering before merge
When not to use it
Not a coding-agent runtime and not a sandbox. If you want the agent to scaffold, generate boilerplate, or explore several implementations, Ponytail will fight you. The first-party bench also says the cut is near zero on code that is already minimal.
FAQ
Does Ponytail cost anything?
No. The site and GitHub list it as a free MIT-licensed plugin. There is no paid tier on ponytail.dev.
Which agents does it work with?
The install block on ponytail.dev covers Claude Code, Codex, Copilot CLI, Gemini CLI and the pi harness, and says it also works with OpenCode, Cursor, Windsurf, Cline, Kiro, Zed and more. The README badge says 20 agents. Check the README for the current list.
Does it delete error handling to save lines?
The site says validation, error handling, security and accessibility are never simplified away. The README's agentic bench reports 100% safety kept vs a bare 'write one-liners' prompt that dropped one guard.
How much less code does it actually write?
The first-party agentic bench claims a mean ~54% less code across 12 feature tasks (Haiku 4.5, n=4) against the same Claude Code agent with no skill, up to 94% on an over-built date picker, and near zero on already-minimal code. That is a vendor self-bench, not an independent study. An older 80–94% single-shot figure is withdrawn in the README.
Share this launch
Embed this badge
<a href="https://orangebot.ai/product/ponytail" target="_blank" rel="noopener noreferrer"> <img src="https://orangebot.ai/api/badge/ponytail.svg" alt="Featured on OrangeBot" width="200" height="54" /> </a>