ponytail makes the agent climb a 7-rung ladder before writing code
Reuse, the standard library and the platform come before new code. Its benchmark shows about half the code at the same correctness, with clear caveats.
Reported by @juliandierker · Verified 2026-11-21
+1 cache this week · +22 GitHub stars since the last radar
- Agent
- Claude Code, Codex
- Stack
- Any
- Skill version
- 5.3.3
Problem
Asked for a date picker, an agent installs a library or hand-builds a calendar.
Ponytail's README counts 335 lines for that, where the repo already had an
Input component and every browser has a native date picker: 10 lines.
Setup
/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail
The ladder runs after the agent has read the code the change touches. It never cuts validation at trust boundaries, data-loss handling, security or accessibility.
What happened
The maintainers' benchmark: 39 tasks, 5 runs each, Opus 5.5 in Claude Code, with and without the skill.
code -53%
output tokens -45%
cost -26%
time -41%
risky logic with a test 98% vs 68%
hidden checks passed 87 vs 86 of 90Sources: README and benchmark.
Reproduce
Install the plugin.
Ask for a feature your repo or platform mostly has already.
Compare the line count and dependencies with a run where you turn it off with
/ponytail off.