# Claude Skills: Anthropic Just Told You Exactly How to Stop Wasting Tokens
**作者**: Mr. Buzzoni
**日期**: 2026-04-15T07:32:37.000Z
**来源**: [https://x.com/polydao/status/2077700506764402788](https://x.com/polydao/status/2077700506764402788)
---
> Anthropic just revealed the blueprint to turn a $20 chatbot into a permanent, specialized engineering team that never forgets your rules
While most people are still wasting 50% of their tokens re-explaining their project context every single morning, a small group of "Skill Architects" just cut their message count by 80% and their token bill in half.
The math is simple:
‣ Without a Skill: 15 messages, 12,000 tokens, 3 failed API calls.
‣ With a Skill: 2 messages, 6,000 tokens, 0 failed calls.
Here is the complete playbook on how to build them, distilled from the official guide into what actually matters for builders
## The Whole Thing Is Just a Folder
A Skill is a directory with one required file and three optional subfolders
That's it. What makes Skills feel heavier than they are comes down to one mechanism
## Progressive Disclosure: The Reason Skills Stay Cheap
Skills load in three levels, so you get specialized behavior without paying for it in tokens on every message.
- Level 1 - YAML frontmatter. Always sits in Claude's system prompt. Tiny. Says only what the skill is and when to reach for it.
- Level 2 - SKILL.md body. Loads when Claude decides the skill fits the task. Full instructions live here.
- Level 3 - linked files in references/. Loads only at the moment it's needed. Big API docs and long guides go here, not in the body.
Two more properties matter. Skills compose, so several can run at once and yours should never assume it owns the whole context. Skills are portable, working identically across Claude.ai, Claude Code, and the API with zero changes
## If You Run an MCP Server, This Is Where It Clicks
Think of MCP as the kitchen: tools, data, equipment. The Skill is the recipe
Connect a server with no skill and users guess the workflow, file support tickets, and prompt differently every time. That inconsistency is the real problem a Skill removes
## The Description Field Decides Everything
This one field is how Claude picks a Skill out of the pile. It has to carry two things: what the skill does, and the exact phrases a user would type to trigger it
Keep it under 1024 characters. Skip angle brackets entirely, since the frontmatter lands in the system prompt and stray markup can inject instructions.
## Naming Rules That Cause Upload Failures
- File must be named exactly SKILL.md (case sensitive). skill.md and SKILL.MD both fail.
- Folder is kebab-case only: notion-project-setup works, Notion_Project_Setup does not.
- No README.md inside the skill folder. Everything goes in SKILL.md or references/.
- Names with "claude" or "anthropic" are reserved and rejected.
## Testing: Go Narrow Before You Go Wide
The method that works: take one hard task, iterate until Claude nails it, then extract the winning version into a skill. That gives faster signal than writing twenty test cases up front. Once it holds, expand across three checks
A real before/after from the guide:
## The Two Failure Modes, and Their Fixes
Undertriggering (never loads): description is too thin. Add trigger phrases and technical keywords. Debug by asking Claude "when would you use this skill?" It quotes the description back and you patch what's missing.
Overtriggering (fires on the wrong things): add negative triggers like "Do NOT use for simple data exploration," and narrow the scope.
## Five Patterns Worth Stealing
One rule threads through all five: for anything that has to be correct, use a bundled script for the check. Code is deterministic. Language interpretation drifts.
## Distribution
- Individuals: zip the folder, upload via Settings > Capabilities > Skills, or drop it in the Claude Code skills directory.
- Teams: admins push skills workspace-wide with automatic updates.
- Programmatic: the API gives a /v1/skills endpoint and a container.skills parameter, running on the Code Execution tool. Best for apps, agents, and pipelines.
## Positioning Matters Too
When you write about your skill, sell the outcome, not the mechanics.
- Good: "Sets up a complete project workspace in seconds instead of 30 minutes of manual setup."
- Bad: "A folder containing YAML frontmatter that calls our MCP tools."
## The Trade Underneath the Whole Guide
You spend thirty minutes encoding a workflow once, and every future session starts from your best version instead of a blank slate. The builders who set this up early stop paying the re-explaining tax on every new chat.
> The full guide and the public skills repo are on Anthropic's site and at github.com/anthropics/skills.
> The skill-creator will scaffold your first one and flag the weak spots before you ship. This is probably just the beginning for how teams package what they know
If you want to see this framework applied at scale, my other piece, "Top 67 Claude Skills That Turn a $20 Subscription Into a Full Dev Team," walks through the actual repos and skill folders builders are using right now. Worth a read once you're ready to go from theory to a stacked skills folder.
> **Mr. Buzzoni@polydao**: [原文链接](https://x.com/polydao/status/2044317956893471081)
>
## And if you found this useful:
- Bookmark this article. The links change and new repos pop up weekly, you'll need this as a reference
- Like and Repost the thread above to help other builders escape the chatbot trap
- For weekly deep dives into AI architecture, quant trading, and the agent economy, follow me: @polydao
- Join the TG Channel: Buzzoni Notes - here I share my raw prompts, custom skills, and alpha that's too early for X
## 相关链接
- [Mr. Buzzoni](https://x.com/polydao)
- [@polydao](https://x.com/polydao)
- [19K](https://x.com/polydao/status/2077700506764402788/analytics)
- [Claude.ai](https://claude.ai/)
- [github.com/anthropics/skills](https://github.com/anthropics/skills)
- [Apr 15](https://x.com/polydao/status/2044317956893471081)
- [4.6M](https://x.com/polydao/status/2044317956893471081/analytics)
- [@polydao](https://x.com/@polydao)
- [Buzzoni Notes](https://t.me/+Wf8q84QkpyJhNjIy)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [6:22 PM · Jul 16, 2026](https://x.com/polydao/status/2077700506764402788)
- [19.9K Views](https://x.com/polydao/status/2077700506764402788/analytics)
- [View quotes](https://x.com/polydao/status/2077700506764402788/quotes)
---
*导出时间: 2026/7/17 13:36:44*
---
## 中文翻译
# Claude Skills: Anthropic 刚刚告诉你如何停止浪费 Token
**作者**: Mr. Buzzoni
**日期**: 2026-04-15T07:32:37.000Z
**来源**: [https://x.com/polydao/status/2077700506764402788](https://x.com/polydao/status/2077700506764402788)
---
> Anthropic 刚刚揭晓了蓝图，能将一个 20 美元的聊天机器人转变为一个永远不会忘记你规则的、永久的专业工程团队
当大多数人仍然每天早上浪费 50% 的 token 重新解释项目上下文时，一小群“技能架构师”刚刚将他们的消息数量减少了 80%，并将 token 账单削减了一半。
数学很简单：
‣ 不使用技能：15 条消息，12,000 个 token，3 次 API 调用失败。
‣ 使用技能：2 条消息，6,000 个 token，0 次调用失败。
以下是如何构建它们的完整指南，这是从官方指南中提炼出的真正对构建者有用的内容
## 整个东西不过是一个文件夹
一个技能就是一个包含一个必需文件和三个可选子文件夹的目录
就这么简单。之所以让技能看起来比实际复杂，归咎于一种机制
## 渐进式披露：技能保持低成本的原因
技能分三个级别加载，因此你无需在每条消息上花费 token 就能获得专业化的行为。
- 级别 1 - YAML 前置元数据。始终位于 Claude 的系统提示词中。非常小。仅说明技能是什么以及何时使用它。
- 级别 2 - SKILL.md 正文。当 Claude 判定该技能适合任务时加载。完整的指令位于此处。
- 级别 3 - references/ 中的链接文件。仅在需要时加载。大型 API 文档和长指南放在这里，而不是正文中。
还有两个属性很重要。技能是可组合的，因此多个技能可以同时运行，你的技能永远不应假设它拥有整个上下文。技能是可移植的，在 Claude.ai、Claude Code 和 API 上的工作方式完全相同，无需任何更改
## 如果你运行 MCP 服务器，这就是恍然大悟的地方
把 MCP 想象成厨房：工具、数据、设备。技能就是食谱
连接一个没有技能的服务器，用户只能猜测工作流程，提交支持工单，并且每次提示都不同。这种不一致性正是技能要解决的真正问题
## 描述字段决定一切
这一个字段就是 Claude 从一堆技能中挑出一个的方式。它必须包含两样东西：技能的作用，以及用户会输入来触发它的确切短语
将其保持在 1024 个字符以内。完全跳过尖括号，因为前置元数据会进入系统提示词，杂乱的标记可能会注入指令。
## 导致上传失败的命名规则
- 文件必须命名为 SKILL.md（区分大小写）。skill.md 和 SKILL.MD 都会失败。
- 文件夹只能使用短横线命名法：notion-project-setup 可以，Notion_Project_Setup 不行。
- 技能文件夹内不能有 README.md。所有内容都放在 SKILL.md 或 references/ 中。
- 包含“claude”或“anthropic”的名称被保留，会被拒绝。
## 测试：先窄后宽
有效的方法是：选一个难题，反复迭代直到 Claude 完美解决，然后将获胜版本提取为一个技能。这比预先编写二十个测试用例能更快获得反馈。一旦稳定下来，再扩展到三个检查
指南中的一个真实的前后对比：
## 两种失败模式及其修复方法
触发不足（从不加载）：描述太单薄。添加触发短语和技术关键词。通过询问 Claude “你什么时候会使用这个技能？”来进行调试。它会引用描述内容，你补上缺失的部分。
过度触发（在错误的情况下触发）：添加负面触发词，如“不要用于简单的数据探索”，并缩小范围。
## 五种值得借鉴的模式
一条规则贯穿这五种模式：对于任何必须正确的东西，使用捆绑脚本进行检查。代码是确定性的。语言解释会产生偏差。
## 分发
- 个人：压缩文件夹，通过 设置 > 能力 > 技能 上传，或将其放入 Claude Code 技能目录中。
- 团队：管理员在工作区范围内推送技能，支持自动更新。
- 编程方式：API 提供 /v1/skills 端点和 container.skills 参数，运行在代码执行工具上。最适合应用、代理和管道。
## 定位也很重要
当你撰写关于你的技能时，要推销结果，而不是机制。
- 好：“几秒钟内设置完整的项目工作区，而不是 30 分钟的手动设置。”
- 坏：“一个包含 YAML 前置元数据的文件夹，用于调用我们的 MCP 工具。”
## 整个指南背后的交易
你花三十分钟一次编码工作流程，未来的每次会话都从你的最佳版本开始，而不是从零开始。早期建立此机制的构建者不再为每次新聊天支付“重复解释税”。
> 完整指南和公共技能库位于 Anthropic 网站和 github.com/anthropics/skills。
> skill-creator 将为你的第一个技能搭建脚手架，并在你发布前标记弱点。这可能只是团队打包其知识方式的开始
如果你想看到这个框架的大规模应用，我的另一篇文章“前 67 个将 20 美元订阅转变为完整开发团队的 Claude 技能”将深入探讨构建者目前正在使用的实际代码库和技能文件夹。当你准备好从理论转向一个丰富的技能文件夹时，值得一读。
> **Mr. Buzzoni@polydao**: [原文链接](https://x.com/polydao/status/2044317956893471081)
>
## 如果你觉得这有用：
- 收藏这篇文章。链接会变，新代码库每周都会出现，你需要将其作为参考
- 点赞并转发上面的帖子，帮助其他构建者逃离聊天机器人陷阱
- 每周深入了解 AI 架构、量化交易和代理经济，请关注我：@polydao
- 加入 TG 频道：Buzzoni Notes - 在这里我分享我的原始提示词、自定义技能以及对于 X 来说太早期的 alpha 信息
## 相关链接
- [Mr. Buzzoni](https://x.com/polydao)
- [@polydao](https://x.com/polydao)
- [19K](https://x.com/polydao/status/2077700506764402788/analytics)
- [Claude.ai](https://claude.ai/)
- [github.com/anthropics/skills](https://github.com/anthropics/skills)
- [Apr 15](https://x.com/polydao/status/2044317956893471081)
- [4.6M](https://x.com/polydao/status/2044317956893471081/analytics)
- [@polydao](https://x.com/@polydao)
- [Buzzoni Notes](https://t.me/+Wf8q84QkpyJhNjIy)
- [升级到 Premium](https://x.com/i/premium_sign_up)
- [6:22 PM · Jul 16, 2026](https://x.com/polydao/status/2077700506764402788)
- [19.9K Views](https://x.com/polydao/status/2077700506764402788/analytics)
- [View quotes](https://x.com/polydao/status/2077700506764402788/quotes)
---
*导出时间: 2026/7/17 13:36:44*
