Ponytail：把「最懒的资深工程师」塞进你的 AI Agent——54% 代码减少，100% 安全

by Mycelium Protocol

---

你认识这个人。马尾辫，椭圆形眼镜，在这家公司待的时间比版本控制系统还长。你给他看五十行代码，他什么都不说，用一行替换了它们。

Ponytail 把他塞进了你的 AI Agent。

97K stars，MIT，2026 年 6 月上线，两个月成为今年 GitHub 增长最快的开发工具之一。

---

问题是什么

你让 Agent 做一个日期选择器。

Agent 安装了 flatpickr，写了一个 wrapper 组件，加了一个样式表，还开始讨论时区问题。

用了 Ponytail 之后：

<!-- ponytail: browser has one -->
<input type="date">

就这一行。浏览器自带了。

这不是刻意偷懒，而是 Agent 在接到任务后先停下来问了一个更基本的问题：这个东西需要被写吗？

---

七级简化阶梯

Ponytail 在 Agent 动手写代码之前，让它先依次过一遍七个问题：

1. 这个东西需要存在吗？   → 不需要：跳过（YAGNI）
2. 代码库里已经有了吗？  → 复用，不要重写
3. 标准库能做吗？         → 用标准库
4. 平台原生特性能做吗？  → 用原生特性
5. 已安装的依赖能做吗？  → 用现有依赖
6. 一行能写完吗？         → 只写一行
7. 以上都不行：最小化实现

在第一个能成立的梯级停下来。

关键是：阶梯运行在理解问题之后，而不是代替理解问题。Ponytail 要求 Agent 在选梯级之前认真读懂涉及的代码、追踪真实调用链。对方案懒，从不对阅读懒。

什么不能省：信任边界的验证、数据丢失处理、安全检查、无障碍（accessibility）——这些从来不在被省略之列。

---

真实基准测试数据

这不是一个 prompt 的单次演示。基准测试的设计是：用 headless Claude Code 编辑一个真实开源仓库（tiangolo/full-stack-fastapi-template，FastAPI + React），12 个功能 ticket，有无 Ponytail 对比，Haiku 4.5，每个 ticket 跑 4 次，看留下的 git diff。

对比基准（无 skill）	代码量	Token	成本	时间	安全
ponytail	-54%	-22%	-20%	-27%	100%
caveman（对照组）	-20%	+7%	+3%	+2%	100%
“YAGNI + one-liners” 纯 prompt	-33%	-14%	-21%	-30%	95%

三个发现：

1. 唯一在所有维度上都下降的方案。 caveman 减少了代码量但增加了 token 消耗；纯 prompt 方式降低了成本和时间，但安全检查只有 95%（它省掉了一个安全防护）。Ponytail 是唯一全部下降且安全 100% 的。

2. 减少幅度与过度构建程度正相关。 日期选择器：从 404 行缩减到 23 行（代码减少 94%）；颜色选择器：从 287 行到 23 行。已经很精简的代码几乎不变。

3. “最少 token”不是目标。 规则是”只写任务需要的”——代码小是结果，不是被高尔夫式削减的产物。

---

与 caveman 的关系

Ponytail 经常被和 caveman 对比，两者可以共存：

• caveman 缩减 Agent 说的话（输出文字）——对代码字节级不动
• Ponytail 缩减 Agent 构建的东西（代码量）——对输出文字不动

“简洁地谈论最小化代码。“两者组合，消耗更少，写出更少，两个维度同时优化。

---

四个强度模式

模式	说明
lite	只激活最保守的几个梯级
full	默认模式，完整七级阶梯
ultra	”当代码库曾经伤害过你个人”
off	关闭

用 PONYTAIL_DEFAULT_MODE 环境变量或 ~/.config/ponytail/config.json 设置每次会话的默认模式。切换命令：/ponytail lite|full|ultra|off。

---

六个内置命令

命令	作用
/ponytail [lite|full|ultra|off]	切换强度 / 查看当前级别
/ponytail-review	审查当前 diff，输出过度工程的删除清单
/ponytail-audit	审查整个仓库（不只是 diff）的过度工程
/ponytail-debt	整理所有被标注为 ponytail: 的技术债备忘
/ponytail-gain	显示基准测试的实测收益（代码减少、成本降低、速度提升）
/ponytail-help	命令快速参考

---

安装方式（20+ 平台）

Claude Code

/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail

（两条命令需要分开发送）

Codex

codex plugin marketplace add DietrichGebert/ponytail
codex plugin add ponytail@ponytail

Gemini CLI / Antigravity CLI

gemini extensions install https://github.com/DietrichGebert/ponytail
# 或 Antigravity：
agy plugin install https://github.com/DietrichGebert/ponytail

OpenCode

{ "plugin": ["@dietrichgebert/ponytail"] }

Cursor / Windsurf / Cline / Kiro / Zed / Aider

从仓库复制对应规则文件：.cursor/rules/、.windsurf/rules/、.clinerules/、.kiro/steering/、AGENTS.md（Jules、Amp、CodeWhale 直接读这个，零配置）。

Subagent 传播：活跃期间，规则集会被自动注入每个通过 Agent 工具生成的子 Agent。用 PONYTAIL_SUBAGENT_MATCHER 环境变量（正则）控制哪些子 Agent 类型接收注入。

---

为什么值得关注

Agent 工具的竞争大多集中在”更智能的单步执行”上：更准的代码生成、更聪明的工具调用。Ponytail 关注的是另一个问题：AI Agent 系统性地过度构建。

这不是模型能力的问题，而是激励结构的问题。Agent 被训练来”完成任务”，完成的证明往往是”写了东西”。没有人告诉它”浏览器自带了日期选择器”比”安装 flatpickr 并写 wrapper”更好——直到 Ponytail 来做这件事。

两个月近 10 万 stars，说明这个问题触到了很多开发者的痛点。

已有多个衍生项目：ponytail-lite（只用一个 AGENTS.md 不依赖插件体系）、ponytail-hermes、ponystack（结合 gstack 流程），以及将 Karpathy LLM 编码指南与 Ponytail 懒惰阶梯合并的实验。

官网（waitlist 开放中）：ponytail.dev
仓库：github.com/DietrichGebert/ponytail

---

“你没写的代码可以无限扩展。零 bug，零 CVE，从来没有宕机过。”

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

Ponytail: The Laziest Senior Dev Inside Your AI Agent

by Mycelium Protocol

You know him. Long ponytail. Oval glasses. He’s been at the company longer than version control. You show him fifty lines. He says nothing. He replaces them with one.

Ponytail puts him inside your AI agent. 97K stars, MIT, launched June 2026 — one of the fastest-growing developer tools on GitHub this year.

---

The Problem

You ask your agent for a date picker.

It installs flatpickr, writes a wrapper component, adds a stylesheet, and starts a discussion about timezones.

With Ponytail:

<!-- ponytail: browser has one -->
<input type="date">

One line. The browser already had one.

This isn’t lazy in the bad sense — the agent read the task carefully, traced the codebase, and asked a more fundamental question: does this thing need to be written at all?

---

The Seven-Rung Laziness Ladder

Before writing code, Ponytail makes the agent stop at the first rung that holds:

1. Does this need to exist?   → no: skip it (YAGNI)
2. Already in this codebase?  → reuse it, don't rewrite
3. Stdlib does it?            → use it
4. Native platform feature?   → use it
5. Installed dependency?      → use it
6. One line?                  → one line
7. Only then: the minimum that works

The ladder runs after the agent understands the problem — not instead of it. Ponytail requires the agent to read the code it touches and trace the real call flow before picking a rung. Lazy about the solution, never about reading.

Always kept: trust-boundary validation, data-loss handling, security, accessibility — never on the chopping block.

---

Real Benchmark Numbers

Not a single-shot demo. The measurement: a headless Claude Code session editing tiangolo/full-stack-fastapi-template (a real FastAPI + React repo), 12 feature tickets, with and without the skill, Haiku 4.5, n=4, scored on the git diff left behind.

vs no-skill baseline	LOC	tokens	cost	time	safe
ponytail	-54%	-22%	-20%	-27%	100%
caveman (terse-prose control)	-20%	+7%	+3%	+2%	100%
“YAGNI + one-liners” prompt	-33%	-14%	-21%	-30%	95%

Three findings:

Only arm that cuts every metric. caveman reduces LOC but increases token consumption; the raw prompt drops cost and time but only 95% safe (it drops a safety guard). Ponytail is the only one that cuts all four while staying fully safe.

Reduction correlates with how much the agent would over-build. Date picker: 404 → 23 lines (94% reduction). Color picker: 287 → 23 lines. Code that was already minimal barely changes.

“Fewest tokens” is not the rule. The rule is “write only what the task needs.” Small code is the outcome, not the target.

---

Relationship with caveman

Ponytail and caveman are complementary, not competing:

• caveman shrinks what the agent says — leaves code byte-for-byte exact
• ponytail shrinks what the agent builds — leaves output prose untouched

Together: terse talk about minimal code. Both running simultaneously, each doing its half.

---

Four Intensity Modes

Mode	Description
lite	Most conservative rungs only
full	Default — full seven-rung ladder
ultra	”For when the codebase has wronged you personally”
off	Disabled

Set per-session default via PONYTAIL_DEFAULT_MODE env var or ~/.config/ponytail/config.json. Switch mid-session with /ponytail lite|full|ultra|off.

---

Six Built-in Commands

Command	What it does
/ponytail [lite|full|ultra|off]	Switch mode or report current level
/ponytail-review	Review current diff for over-engineering, return a delete-list
/ponytail-audit	Audit the whole repo, not just the diff
/ponytail-debt	Harvest deferred ponytail: shortcuts into a ledger
/ponytail-gain	Show the measured impact scoreboard from the benchmark
/ponytail-help	Quick command reference

---

Install (20+ Platforms)

Claude Code:

/plugin marketplace add DietrichGebert/ponytail
/plugin install ponytail@ponytail

(Two separate prompts required)

Codex:

codex plugin marketplace add DietrichGebert/ponytail
codex plugin add ponytail@ponytail

Gemini CLI / Antigravity CLI:

gemini extensions install https://github.com/DietrichGebert/ponytail

OpenCode:

{ "plugin": ["@dietrichgebert/ponytail"] }

Cursor / Windsurf / Cline / Kiro / Aider: copy the matching rules file (.cursor/rules/, .kiro/steering/ponytail.md, etc.). Agents that auto-load AGENTS.md (Jules, Amp, CodeWhale, Swival, Qoder) work with zero setup from the repo root.

Subagent propagation: the ruleset is injected into every subagent spawned via the Agent tool. Use PONYTAIL_SUBAGENT_MATCHER (a regex against agent type) to scope the injection.

---

Why It Matters

Competition in agent tooling has focused on smarter single-step execution — better code generation, more accurate tool calls. Ponytail addresses a different problem: AI agents systematically over-build.

This isn’t a model capability problem; it’s an incentive structure problem. Agents are trained to “complete tasks,” and completion is often evidenced by “wrote something.” Nobody told the agent that <input type="date"> is better than flatpickr + wrapper + stylesheet — until Ponytail does.

Nearly 100K stars in two months shows this hits a real nerve. The ecosystem is already branching: ponytail-lite (a single AGENTS.md without the plugin infrastructure), ponystack (combining gstack’s process with Ponytail’s restraint), Karpathy–Ponytail skill fusions, and more.

Waitlist: ponytail.dev · Repository: github.com/DietrichGebert/ponytail

---

“The code you never wrote scales infinitely. Zero bugs, zero CVEs, 100% uptime since forever.”

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.