正在打开每日关注
正在打开每日关注
两个 GitHub 技能分别压缩代理输出与代码规模,主张组合使用并自行基准测试。
the two employees your claude code / codex should hire. both grabbed the #1 spot on github trending weekly at some point. one saves you words, one saves you code. caveman, 81k stars, 𝗺𝗮𝗸𝗲𝘀 𝘆𝗼𝘂𝗿 𝗮𝗴𝗲𝗻𝘁 𝗾𝘂𝗶𝘁 𝘁𝗵𝗲 𝘀𝗺𝗮𝗹𝗹 𝘁𝗮𝗹𝗸. the avatar is a gray rock and the resume is probably carved on one too. answers in 19 tokens: Bug in auth middleware. Fix. done. → output tokens -75% → same react diagnosis: 69 tokens down to 19 → works on claude code, codex, cursor, 30+ agents ponytail, 72k stars, 𝗰𝘂𝗿𝗲𝘀 𝗼𝘃𝗲𝗿-𝗲𝗻𝗴𝗶𝗻𝗲𝗲𝗿𝗶𝗻𝗴. long greasy ponytail, oval glasses, been at the company longer than the version control. you hand him 50 lines, he pushes up his glasses and hands back 1. ask for a date picker and he gives you <input type="date">, won't even install flatpickr. → code -54%, up to -94% when the agent over-builds → cost -20%, speed +27% → measured on 12 real claude code tasks on a real fastapi + react repo i assumed it was pick one. read both repos, they manage different things, you can hire both. ponytail even dragged caveman into its benchmark as a control arm: 20% less code but 𝗰𝗼𝘀𝘁 𝘄𝗲𝗻𝘁 𝘂𝗽 𝟯%. that's ponytail's own test though. worth installing both and running your own round, your repo and your tasks are not theirs
是什么 这条推文推荐两个定位互补的 GitHub Agent 技能。Caveman 负责让编码代理用更少的话表达同一技术判断,README 以 69 个 token 对比 19 个 token,并声称可减少约 75% 输出。Ponytail 负责抑制过度工程化,强调少写代码、少装依赖;日期选择器直接使用原生 input 即是典型例子。
核心设计 从 README 能确认的架构边界是:二者都以 skill/plugin 形式接入 Claude Code、Codex 等多个编码代理,但没有给出运行时实现、提示词注入方式或内部模块结构。它们的设计分工清晰:Caveman 优化沟通带宽,Ponytail 优化实现范围,组合后分别控制代理说什么和写多少。
关键洞察 这不是简单的二选一,而是把 Agent 效率拆成输出成本与代码复杂度两个变量。Ponytail 页面自述在真实 FastAPI + React 仓库、12 个任务中平均少写约 54%,并伴随成本下降与速度提升;这种按任务观察的 benchmark 比单一示例更有参考价值。
边界与风险 所有指标、星标数和趋势榜单都来自项目自述,不能直接视为普适结论。推文也提到 Ponytail 自测中 Caveman 少写约 20% 代码但成本上升约 3%。应在自己的仓库、模型和任务集上复测,并重点检查安全护栏、可维护性与遗漏需求。画面主要是 README 展示,视觉包装服务于工程传播,不构成设计方法本身。