2026年7月16日 · Skills · Knowledge System

我开始用 find-skills 管理 AI 能力，而不是死记 Prompt

Prompt 收藏夹越大越难用。我更需要的是一个能力索引：遇到任务时，先找到合适的工作流，而不是翻收藏夹。

痛点

我以前管理 prompt 的方式很原始：看到一段有用的话就存起来，起个名字，放进某个文件夹。刚开始很爽，十几个 prompt 的时候确实够用。到几十个以后，问题来了：名字记不住，场景分不清，相似模板重复，真正要用的时候反而先在收藏夹里迷路。

find-skills 解决的是另一层问题

find-skills 解决的不是"帮我写一句更好的 prompt"，而是另一个更底层的问题：我现在到底需要哪一种能力？

任务	我需要的能力
写博客	blog-writing-guide
找最近讨论	last30days
做决策拷问	grilling

任务不同，应该触发的工作流不同。find-skills 在 skills.sh 上是安装量第一的 skill（2.7M installs），说明这不是我一个人的痛点。

重新理解个人 AI 工具箱

这让我开始重新理解个人 AI 工具箱。

Prompt 是一句指令，skill 是一套操作规程。

一个好的 skill 通常会写清楚什么时候用、输入是什么、步骤是什么、输出应该长什么样、什么时候不要用。它不是魔法咒语，更像一本小型 SOP。

价值：从"记得有"到"能搜到"

find-skills 的价值在于，它把"我记得有个东西能干这个"变成"我可以搜索并验证有没有合适的能力"。这一步很重要，因为 AI 能力越来越多之后，真正的瓶颈不是生成，而是调度。你要先知道该调用哪套知识，再谈怎么生成结果。

我现在整理 AI Shelf，也准备按这个逻辑来：每个条目不只写名字和链接，而要写使用场景、输入要求、输出形态、适用边界和我的使用案例。这样别人看到的不是一堆工具名，而是一套可以复用的工作方式。

适用边界

场景	是否该用 find-skills
任务方向没定、不知道该用哪个能力	✅ 先搜能力，再动手
已经知道要用哪个 skill	❌ 直接装，不用搜
低质量 skill 包装的糟糕 prompt	⚠️ 搜到也要先读、小范围试用再沉淀

低质量 skill 只会把一个糟糕 prompt 包装成更正式的糟糕 prompt。我的做法是，先找，再读，再小范围试用，最后才沉淀到自己的方法库里。

I Use find-skills to Manage AI Capabilities, Not Memorize Prompts

The more prompts you save, the harder it becomes to use them. What I really need is a capability index: when a task comes in, find the right workflow first—don't dig through the folder.

The Pain Point

My old approach to managing prompts was primitive: whenever I found something useful, I'd save it, name it, and file it. It worked fine with a dozen prompts. But once I hit dozens, the problems piled up: I couldn't remember names, couldn't distinguish use cases, had duplicate templates, and ended up getting lost in my own collection when I actually needed something.

What find-skills Actually Solves

find-skills doesn't solve "write me a better prompt"—it addresses a more fundamental question: what capability do I actually need right now?

Task	Capability I Need
Write a blog post	blog-writing-guide
Find recent discussions	last30days
Pressure-test a decision	grilling

Different tasks should trigger different workflows. find-skills is the #1 most-installed skill on skills.sh (2.7M installs)—clearly this isn't just my problem.

Rethinking the Personal AI Toolkit

This made me reconsider how I organize my personal AI toolkit.

A prompt is an instruction. A skill is a standard operating procedure.

A good skill typically specifies: when to use it, what inputs it expects, what steps to follow, what the output should look like, and when NOT to use it. It's not a magic incantation—it's more like a mini SOP.

Value: From "I Remember Having Something" to "I Can Find It"

find-skills transforms "I think there's something for this" into "I can search and verify whether the right capability exists." This matters because as AI capabilities multiply, the real bottleneck isn't generation anymore—it's orchestration. You need to know which knowledge set to invoke before you can talk about generating results.

I'm now organizing AI Shelf the same way: each entry won't just have a name and link—it will include use cases, input requirements, output format, scope boundaries, and my personal usage notes. So readers see not a list of tools, but a set of reusable workflows.

Scope of Use

Scenario	Should you use find-skills?
Direction unclear, unsure which capability to use	✅ Search for capability first, then act
Already know which skill to use	❌ Install directly, no need to search
Low-quality skill wrapping a bad prompt	⚠️ Read first, test small, then adopt

A low-quality skill just wraps a bad prompt in a more official package. My approach: search, read, test in a small scope, then add to my personal toolkit.