Ponytail

An AI agent skill that makes a coding agent “think like the laziest senior dev in the room” — “the best code is the code you never wrote.” Where most of the spoke’s skills add capability (agentic-seo-skill, claude-skills-ppc, pm-skills), Ponytail is a skill of restraint: it constrains the agent to write only what a task genuinely needs, attacking the over-engineering failure mode where agents reach for a new dependency, a wrapper component, or an abstraction when a simpler path exists. First-party project repo (DietrichGebert, MIT, ~45.9k★ at snapshot). URL-only ingest.

The mechanism — a decision ladder

The skill’s core is a priority ladder the agent walks before writing code: Does this need to exist? → Does stdlib do it? → A native platform feature? → An already-installed dependency? → Can it be one line? → Then the minimum that works. Each rung prefers doing less; new code is the last resort, a new dependency nearly the last. It ships intensity modes — lite, full, ultra, off — so the restraint can be dialed up or down, plus commands to review a diff, audit a repo, “harvest” deferred shortcuts, show benchmarks, and get help. A stated guardrail: restraint never trades away error handling, validation, security, or accessibility — it cuts abstraction, not correctness.

Why it belongs here, and what it adds

• A new flavor of skill: restraint, not capability. The spoke has the augment pole (lathe, renwei-writing) that keeps the human in the loop; Ponytail is a different counter-current — it keeps the agent from doing too much. The shared shape is that a loadable skill is pole-agnostic: the same progressive-disclosure primitive can add expertise or impose discipline on output. Ponytail makes “less” the deliverable.
• A strong standardization-reach data point. It advertises support across an unusually wide host roster — Claude Code, Codex, GitHub Copilot CLI, OpenCode, Gemini, Antigravity, CodeWhale, OpenClaw, plus editor extensions (Cursor, Windsurf, Cline, Kiro, Zed, Aider) — via plugin marketplaces, rules-file copies, or package managers depending on the host. One skill targeting 15+ agents is direct evidence for the portable-skill thesis, though (as with renwei-writing‘s Cola dir) the delivery is still per-host rather than one universal format.
• It connects to the verification thread. Its review/audit commands are output-grounded checks (ai-job-search, renwei-writing discipline) pointed at a different target — flagging over-engineered diffs instead of layout or voice.

Measured independently, 2026-07-28 — the effect is real and about a third the claim

jetbrains-ponytail-tested is the outside test the caveat below was waiting for: 80 paired SkillsBench tasks, claude-sonnet-5, 251 billed trials, with per-trial audits confirming the ruleset reached the treatment arm and never the baseline.

Code −15.4% (p=0.088), cost −10.3% (p=0.004), wall-clock −11%, and no quality difference detected (9 tasks worse, 6 better, 65 identical). Against the advertised −54% code / −20% cost, the measured effect is roughly a third — explained, fairly, by method rather than dishonesty: the repo’s figure is a mean over twelve hand-picked feature tickets, JetBrains’ is a median over 80 tasks chosen by nobody for this purpose.

The benefit is conditional. Savings concentrate where the baseline over-builds (code fell 31% on larger builds) and approach zero on already-lean tasks. Ponytail doesn’t make an agent write less; it stops an agent writing too much, so the payout tracks how often the baseline was over-engineering in the first place. Any single headline percentage is an artifact of the task mix.

Two adherence findings land on the skill rather than the benchmark. Without the SessionStart hook it self-activated zero times in ten sessions, despite advertising readiness for “ANY coding task” — so the measured benefit depends on the always-on install. And its own rule about leaving ponytail: comments on deliberate shortcuts was followed once in 80 trials.

What the test cannot speak to is the guardrail below: SkillsBench scores task completion, not security, validation, error handling or accessibility. The claim Ponytail makes about what it won’t trade away remains unmeasured.

Caveat — the numbers are self-reported

The headline claims (~54% less generated code on average, up to 94% in over-engineering cases; 20% cost reduction, 27% faster execution, 100% safety maintained) are the repo’s own benchmarks, with no published methodology or independent replication. They land squarely in the spoke’s standing benchmark open question: a README metric, not a neutral measurement. Recorded as the project’s claim — and, as of 2026-07-28, partly borne out and substantially overstated (above). Author (DietrichGebert) noted inline rather than as an entity node — pseudonymous GitHub handle, low graph signal, matching the skill-pack precedent.

The contrast arrived 2026-07-28. i-have-adhd is the same species — a restraint skill, MIT, cross-harness, aimed at over-production — but it constrains the agent’s output where Ponytail constrains its code, and it inverts the evidence posture: no headline metric, and instead a shipped evals/ harness with a blind weighted rubric and a release gate that a change must clear (correctness and safety within 0.1 of baseline; concision worth only 10%). Ponytail’s guardrail — restraint never trades away validation, security, accessibility — is the same commitment i-have-adhd encodes as a scored threshold rather than a promise in prose. One published a number without a method, the other a method without a number — and later the same day jetbrains-ponytail-tested supplied the third posture neither had: a number produced by someone with nothing to sell. i-have-adhd’s harness is still unrun. See reasoning-scaffolds for where both sit as output-shape skills.

Related

agent-skills · agentskills-spec · lathe · renwei-writing · ai-job-search · i-have-adhd · reasoning-scaffolds · agent-evals · jetbrains-ponytail-tested · synthesis