Ponytail Github

Ponytail Github

Ponytail Github is an opinionated prompt-engineering skill for AI coding agents, framed as the persona of a veteran engineer who has been quietly deleting code since before version control existed. The skill installs a six-step ladder into the agent — does this need to exist, does the standard library do it, does the platform do it, does an installed dependency do it, can it be one line, then the minimum that works — and refuses to cut trust-boundary validation, error handling, security, or accessibility under any circumstance.

Technically the project is a single GitHub repository at DietrichGebert/ponytail: a markdown skill file, two tiny Node.js lifecycle hooks (under a hundred lines combined), a Spanish translation, and a reproducible benchmark suite built on promptfoo and a headless Claude Code harness running against the FastAPI full-stack template. It plugs into 14 popular agents (Claude Code, Codex, Cursor, Aider, Continue, Cline, Roo Code, Windsurf, GitHub Copilot, Gemini CLI, OpenCode, Qwen Code, Cody, Amp) and activates on every prompt with no UI changes.

What separates Ponytail Github from a typical terse-prompt experiment is the rigor of its measurement. The headline figure of 54% fewer lines, 22% fewer tokens, 20% lower cost and 27% faster runs is the mean across twelve real feature tickets on a real FastAPI plus React codebase, not a single cherry-picked prompt. On over-build traps the reduction reaches 94%, while on already-minimal code it is near zero — exactly the curve you would hope for from a real minimalism skill, not a code-golf trick. Every claim is backed by per-task tables and reproducible scripts in the benchmarks/ directory, and the project is MIT licensed so it can be adopted or extended freely.