Agent skill · MITworks with 20 agents
The best code is
the code you
never wrote.
Ponytail makes your AI agent think like the laziest senior dev in the room. The rule was never “fewest tokens” — it is write only what the task needs, and never cut validation, error handling, security, or accessibility. The code ends up small because it is necessary, not golfed.
Before writing code, stop at the first rung that holds
rung 4 wins here- 1Does this need to exist?no → skip it
- 2Already in this codebase?reuse it
- 3Stdlib does it?use it
- 4Native platform feature?use it ◀
- 5Installed dependency?use it
- 6One line?one line
- 7Only then: the minimum that workswrite it
The ladder stops; it does not optimise. Nothing below rung four runs once rung four holds — which is why the date picker never reaches a dependency discussion.
Without the skill
You ask for a date picker.
- Installs flatpickr
- Writes a wrapper component
- Adds a stylesheet
- Starts a discussion about timezones
With ponytail — rung 4 held
<!-- ponytail: browser has one -->
<input type="date">
The project reports this task going from 404 lines to 23, because the ladder reaches a native input before it reaches a component. Both panels are quoted from the README.
Against a no-skill baseline
The project’s own agentic benchmark: a headless Claude Code session editing a real FastAPI + React repository, twelve feature tickets, n=4, Haiku 4.5, scored on the diff it leaves behind. Figures reproduced as published.
The last column is the one worth reading twice: the bare “write one-liners” prompt is the only arm that drops a safety guard. The project also publishes its own correction — an earlier single-shot run reported 80–94%, which it now describes as a per-task ceiling rather than an average, and labels a conversational-baseline artifact.