An AI agent learns the senior engineer's cut: half the code, and risky logic ships with tests more often
Synopsis
Ponytail 5 constrains how an AI agent writes code through a single prompt: write only what the task needs, but never cut validation, error handling, security or accessibility. In Claude Code, the same agent with and without the skill across 39 tasks cut code by 53%, time by 41%, cost by 26% and tokens by 45%; risky logic shipped with a test 98% of the time versus 68% without. All 30 runs on six security tasks passed.
Interpretation
The agent is told to look at what the repository already contains before deciding how much to write, instead of building an implementation from scratch. Without the skill, asked for a date picker, the agent installs a date-picker library or hand-builds a whole calendar: 335 lines. With it, the agent finds an existing Input component and the browser's built-in date picker and combines them into 10 lines; the comparison ran in Claude Code with the same agent with and without the skill, 39 tasks including a real FastAPI + React repo, Opus 5.5, 5 runs each.
The boundary of the saving is written down: trust-boundary validation, data-loss handling, security and accessibility are never on the chopping block. The rule had been easy to read as fewest tokens, while the actual rule is to write only what the task needs. Logic with a branch, a loop, a parser, money or security leaves one small test behind; six security tasks (SQL injection, path traversal, forged tokens, rate limiting, malformed CSV rows, caching) passed all 30 runs.
The companion review and audit commands shift from finding code to delete toward reading the code a change touches the way an on-call senior developer would. The old review looked only for code to cut, and the old audit only listed what to delete. They now check bugs, security, real load, missing tests and speed, and each finding says what the code does, what goes wrong, how to fix it and what happens if you don't; the audit maps entry points, data flow and expected load before ranking what to fix first.
Perspective
The approach targets agent users writing code in Claude Code or Codex and is distributed as one prompt plus a rules file. The benchmark ran in Claude Code with the same agent with and without the skill, 39 tasks including a real FastAPI + React repo, Opus 5.5, 5 runs each. The cost and token band needs Claude Code 2.1.290 or later and shows in the terminal or desktop app, not in the VS Code chat panel, cloud sessions or claude -p.
In the project's own comparison, a blind test had Ponytail 5's replies beating the previous Ponytail's 110 to 67, a comparison of reply preference only. The six security tasks passed all 30 runs, covering those six scenarios. The reductions in code, time, cost and tokens come from the 39-task benchmark setting, and the source does not state whether the same proportions hold on other repositories or task types.