Give Agents Libraries, Not Instructions

Put scripts and helper functions in skill folders. Then agents spend tokens on composition and decision-making instead of rewriting boilerplate.

The insight

Anthropic noticed something specific: “one of the most powerful tools you can give Claude is code.” When a skill ships with executable scripts and function libraries, Claude stops spending tokens reinventing data-fetching code. It just calls what’s already there. The tokens go into the higher-order question instead.

For example, a data science skill includes helpers for fetching events. A user asks: “What happened on Tuesday?” Claude doesn’t write the fetching code. It writes a script that composes the existing functions: filter, aggregate, analyze.

The technique

Find the operations you keep repeating in your workflows: API calls, data transforms, file operations. Pull them into a script library inside the skill folder. Document the library in SKILL.md. The agent finds the functions and composes them instead of reinventing them every session.

The shift is from reconstruction to composition. A much higher-leverage use of both tokens and the context window.

CLI-Anything: the logical extreme

CLI-Anything (HKUDS, 21.5k GitHub stars) pushes this all the way. A 7-phase automated pipeline turns any desktop software into an installable CLI package. After pip install, the agent can drive Blender, GIMP, OBS, and others through structured JSON commands.

The agent never learns the underlying API or GUI. It calls a pre-built CLI library. Each generated CLI ships with a SKILL.md for auto-discovery, a REPL for exploration, and undo/redo for safe state management.

This is the most literal version of “give agents libraries”: the library is the CLI package.

The principle

Any time you catch yourself writing the same kind of instructions to an agent twice, extract a callable library instead. Instructions are consumed once and forgotten. Libraries persist across sessions and compose with each other.

把脚本和辅助函数放进 skill 文件夹。Agent 的 token 就能花在组合和判断上，不用再重写样板代码。

核心洞察

Anthropic 发现了一件挺具体的事：“你能给 Claude 的最强大的工具之一，就是代码。”如果 skill 里带着能执行的脚本和函数库，Claude 就不用花 token 重新生成获取数据的代码了。它直接调已有的函数，把精力放在更高阶的问题上。

举个例子，一个数据科学 skill 自带获取事件的辅助函数。用户问“周二发生了啥？”，Claude 不写数据获取那一段，直接写个脚本，把现成的函数串起来：过滤、聚合、分析。

技巧

把工作流里反复出现的操作识别出来：API 调用、数据转换、文件操作。提取成一个脚本库，放进 skill 文件夹。库的内容在 SKILL.md 里记录一下。Agent 自己会发现并组合，不用每次会话都重新造一遍轮子。

精力就从重建挪到了组合。同样的 token 和上下文窗口，能撬动的东西多了一大截。

CLI-Anything：逻辑极限

CLI-Anything（HKUDS，21.5k GitHub stars）把这个思路推到了头。一条 7 步自动化流水线，把任意桌面软件变成一个可安装的 CLI 包。pip install 完，agent 就能通过结构化 JSON 命令操控 Blender、GIMP、OBS 这些软件。

Agent 不用学软件本身的 API 或者 GUI。它调一个预先构建好的 CLI 库就行。每个生成的 CLI 都自带一份 SKILL.md 供自动发现，一个 REPL 方便探索，还有 undo/redo，状态改坏了能撤回。

这是“给 agent 库而不是指令”最字面的实现：库就是 CLI 包本身。

原则

只要你发现自己在反复给 agent 写同一类指令，就该把它抽出来，做成一个能调用的库。指令用一次就没了。库跨会话留着，还能互相组合。

Give Agents Libraries, Not Instructions给 Agent 库，而非指令