Ponytail
ponytail is a Claude Code skill (npm: @dietrichgebert/ponytail) that enforces minimal code generation in AI coding agents. It installs a fictional "laziest senior dev" persona into the agent's system prompt, instructing it to prefer built-in browser/platform primitives over imported libraries, and to write only what the task strictly requires.
The skill adds a constraint ladder the agent follows before writing any code: reach for what already exists (native HTML, stdlib, existing dependencies) before introducing anything new. It does not reduce safety guards -- error handling, validation, and accessibility requirements are preserved.
Measured on headless Claude Code sessions editing a FastAPI + React open-source repo, 12 feature tickets, Haiku 4.5, n=4:
| Metric | Change vs no-skill baseline |
|---|---|
| Lines of code | -54% (up to -94% on over-build traps) |
| Token usage | -22% |
| Cost | -20% |
| Time | -27% |
| Safety (adversarial tier) | 100% |
The largest cuts occur where the agent would otherwise reach for a third-party component (e.g. a date picker library) when <input type="date"> suffices. On already-minimal code the gain is near zero.
53,778 GitHub stars as of June 2026, two weeks after the June 12, 2026 release. MIT license. Benchmarks are reproducible via promptfoo and documented at benchmarks/results/2026-06-18-agentic.md in the repo.
Placed in Tools/Trial inner. Installed as a Claude Code plugin and used in production on this repo's helper scripts (add_blip.py, update_blip.py): /ponytail-audit identified bloated docstrings and a duplicated JSON loader; findings were applied, cutting 35 lines. The benchmark methodology is sound and reproducible; the 53k-star adoption signal is unusually strong for a two-week-old tool. Inner position reflects confidence in the mechanism and the quality of the evidence.
