boxmoe_header_banner_img

Hello! 欢迎来到阿虪的开发世界~

加载中

文章导读

improve:让最强的模型只做审计写计划,执行交给便宜模型


avatar
阿虪 12 11 月, 2026 194
improve

中文版

improve(GitHub:https://github.com/shadcn/improve ,2026-06-10 创建,今晨抓取时已 10,881★)是 shadcn 写的一个 Agent 技能:审计任何代码库,然后给别的 Agent 写出可执行的实施计划。它自己从不写代码——“The plan is the product.”(计划就是产品)。

README 的思路很直接:把最强的模型用在智能最能复利的地方——理解代码库、判断什么值得做、写出规格说明;执行交给便宜的模型。用法是 npx skills add shadcn/improve 装上,然后 /improve 全量审计并输出按优先级排好的发现与计划,/improve quick 只做快速扫描、/improve deep 穷尽式审计、/improve security 专注安全(还有性能、测试、缺陷等模式)、/improve branch 只审当前分支的改动、/improve next 给出下一步功能建议、/improve plan <描述> 跳过审计直接写某个功能的规格、/improve execute <计划> 派便宜的执行器去干活并复核。计划是纯 Markdown,任何 Agent(或人)都能接手。日榜今天单日新增 361 星。

为什么值得了解

  • 它把“贵模型做决策、便宜模型做执行”变成了可落地的分工:对天天用 AI 写代码的人来说,这是最直接的省钱姿势
  • 计划是产品:审计输出的是一份自包含的规格文档,而不是一堆聊天记录,换个 Agent、换个人都能接着干
  • 作者是 shadcn(shadcn/ui 的作者),技能遵循 Agent Skills 格式,装上就能在任何支持该格式的助手里用

English Version

improve (GitHub: https://github.com/shadcn/improve , created 2026-06-10, 10,881 stars at this morning’s fetch) is an agent skill from shadcn that audits any codebase and writes implementation plans for other agents to execute. It never implements anything itself — “The plan is the product.”

The README’s idea is straightforward: use your most capable model where intelligence compounds — understanding the codebase, judging what’s worth doing, writing the spec — and hand execution to cheaper models. Install it with npx skills add shadcn/improve, then /improve runs a full audit and outputs prioritized findings and plans; /improve quick does a cheap pass, /improve deep an exhaustive one, /improve security focuses on security (with perf, tests and bug modes too), /improve branch audits only what the current branch changes, /improve next suggests where to take the project, /improve plan specs one thing without the audit, and /improve execute dispatches a cheaper executor and reviews its work. Plans are plain Markdown, so any agent (or human) can pick them up. It gained 361 stars today alone.

Why it’s worth knowing

  • It turns “expensive model decides, cheap model executes” into a workable division of labour: for anyone coding with AI daily, this is the most direct way to save money
  • The plan is the product: the audit outputs a self-contained spec document, not a pile of chat logs, so another agent or another person can carry on from it
  • The author is shadcn (of shadcn/ui fame), the skill follows the Agent Skills format, and it works in any assistant that supports that format once installed


评论(0)

查看评论列表

暂无评论


发表评论

表情 颜文字

插入代码