Your agent overbuilds. Every time. A keyless YAGNI gate, measured.

Ask a coding agent for a small feature and it hands you a config layer, two abstractions and a "for later" option nobody asked for. DietrichGebert/ponytail is a skill that makes the agent behave like the laziest senior developer in the room and ship the smallest change that fully solves the task, so we wrapped it in a keyless minilab whose report (run on October 13, 2026) passed 3 of 3 gate rules, verified 1 skill and ran 172 of 172 upstream tests green in 66 s. The catch is that a YAGNI skill only trims what you feed the agent, so a swarm of them can still spend whatever you let it.

The catch

Overbuilding is cheap for one agent. It stops being cheap when a lead agent fans the habit out to a swarm. Anthropic documents up to 64 concurrent threads for multi-agent orchestration, and says that number is not guaranteed (managed agents: multi-agent orchestration). So treat "64× faster" as a ceiling. At full concurrency, scope creep in the prompt becomes scope creep in up to sixty-four workers.

Two more hedges apply. Ponytail's author reports roughly "~53% less code"; this lab does not re-run or certify that figure. And a YAGNI skill is no substitute for a budget, so keep your session budget and kill switch.

What we ran (keyless)

bun install && bun run report

scripts/report.ts runs three steps in order and writes reports/gate.json plus reports/gate.md:

1. gate:skill checks that the ponytail skill exists at .claude/skills/ponytail/SKILL.md and that there is no mirror under .agents/skills or .cursor/skills. Three rules, each printed as a line the report counts.
2. verify:skills-install runs skills@1.7.2 add DietrichGebert/ponytail --skill ponytail -a claude-code --copy into a disposable directory, then asserts the file landed on the canonical path and no .agents mirror appeared.
3. test:upstream clones ponytail at the pinned commit 2961b52fc1fe6a5a87c7a959a79512e5c2f3ffe5, builds a throwaway Python venv with pandas for its CSV probe, then runs upstream npm test (node --test tests/*.test.js && npm test --prefix pi-extension). The report sums the node --test summary lines across both runs.

From reports/gate.json (run on October 13, 2026):

The first pass of bun run report went red. verify:skills-install failed on a missing yaml package inside a stale bunx cache of skills@1.7.2. Clearing the cache and re-running went green with no code change. The script records that kind of failure instead of hiding it. A failed step lands in gate.json with its exit code, and an upstream failure that looks like a network error is flagged as such.

How it is wired

Prompt → Ponytail gate → Plan → Build, with a dashed "rejected scope" edge from the gate back to the prompt. The skill lists every place the change must reach, then takes the first option that fully works: skip features nobody asked for, reuse what the codebase has, prefer the standard library, never add a dependency for a few lines. Rejected scope goes back to the human as a one-line note instead of as code.

The minilab adds what a skill cannot do for itself, which is proof that it is installed where the agent will find it. Skills live only under .claude/skills/, so Claude Code and Cursor discover them on one path, and a mirror under .agents/skills or .cursor/skills fails gate:skill in CI. The pinned upstream run means a future ponytail release cannot silently change what "passes" means; you bump the pin on purpose.

Upstream metadata, from gh api repos/DietrichGebert/ponytail on October 13, 2026: JavaScript, MIT, 162,411 stars.

The tool

ponytail is the YAGNI skill, and its rule is the smallest complete change. In chat, /ponytail full is the default lazy-senior mode, /ponytail lite narrows scope further, and /ponytail ultra pushes back on the request itself. Say "stop ponytail" or "normal mode" to end it.

Minilab (open source)

Install or refresh the committed skill:

DISABLE_TELEMETRY=1 bun --bun x skills@1.7.2 add DietrichGebert/ponytail \
  --skill ponytail -a claude-code -y --copy

Then run bun run report and read reports/gate.md for your own numbers.

bun install && bun run report