Agent Skills 标准捐献与技能编写六条铁律
Agent Skills 标准捐入 Linux Foundation,附官方技能编写最佳实践与一键沉淀技能的提示词。
原始内容
Anthropic just donated the Agent Skills standard to the Linux Foundation, the same home MCP went to. Since it launched last October, 𝟭 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝘀𝗸𝗶𝗹𝗹𝘀 have been published and installed nearly 280 million times, and 70+ agent platforms support it, including ChatGPT, Cursor, GitHub Copilot and VS Code. Over that year, how to write a skill settled into a standard method, and the official guide went to the foundation too. The 6 most useful rules from it: → Do the task with your agent for real first, then turn the steps that worked, the corrections you made, the input and output formats and the context you added into the skill. A model writing from general knowledge gives you filler like "handle errors appropriately" → Ask of every line: would the agent get this wrong without it? If not, cut it → At startup the agent reads only the name and description, so the description decides whether the skill ever gets used. Write it as "use this skill when...", lean pushy, list the cases users won't spell out, and stay under 1,024 characters → Keep SKILL.md under 500 lines and 5,000 tokens, move details into references/, and say when to read which file → Hard-code the exact command for fragile steps, explain the why where there's room for judgment, and when there are several ways to do something, give one default → In many skills the most valuable part is 𝘁𝗵𝗲 𝗴𝗼𝘁𝗰𝗵𝗮𝘀 𝗹𝗶𝘀𝘁: every time you have to correct the agent, add a line Right after your agent finishes a task you hand it often, send this to Claude Code in the same session to save how you did it as a standard skill 👇 "Following https://t.co/YN17Wv07wD, turn the task we just did into a skill. Use only the steps that worked in this conversation, the places I corrected you, and related fixes in CLAUDE.md, AGENTS.md and the git history; don't fill gaps from general knowledge. Write the description as 'use this skill when...', include cases I didn't spell out but that should trigger it, and keep it under 1,024 characters. Keep SKILL.md under 500 lines, move details into references/ with a note on when to read each file, and give the gotchas their own section. Use lowercase letters, numbers and hyphens for the name, matching the folder name. Finally, write 2-3 real test cases in evals/evals.json (prompt plus expected output), run each one, and show me the results and what you did."
We're donating Agent Skills to the Agentic AI Foundation (@AgenticAIFdn), the same foundation we donated MCP to. Since December 2025, 70+ agent platforms have integrated the standard, including ChatGPT, Cursor, GitHub Copilot, and VS Code. Here's what that means for you:
中文译文
机器翻译,保留原文供核对。
Anthropic 刚刚把 Agent Skills 标准捐给了 Linux 基金会,和 MCP 去的是同一个家。自去年 10 月发布以来,已有 100 万个技能被发布、被安装近 2.8 亿次,70+ 个 agent 平台支持它,包括 ChatGPT、Cursor、GitHub Copilot 和 VS Code。 这一年里,「怎么写一个技能」沉淀成了一套标准方法,官方指南也一并交给了基金会。其中最有用的 6 条规则: → 先让你的 agent 真刀真枪把任务做一遍,然后把跑通的步骤、你做过的纠正、输入输出格式和你补充的上下文变成技能。让模型凭通用知识去写,你只会得到「妥善处理错误」这类废话 → 对每一行都问一句:没有它 agent 会做错吗?不会就删掉 → 启动时 agent 只读名字和描述,所以描述决定了这个技能到底会不会被用上。把它写成「use this skill when...」,宁可写得强势一点,把用户不会明说的场景列出来,并保持在 1,024 个字符以内 → SKILL.md 保持在 500 行、5,000 tokens 以内,细节移到 references/,并说明什么时候该读哪个文件 → 脆弱的步骤把确切命令写死,有余地做判断的地方解释为什么,一件事有多种做法时给一个默认做法 → 很多技能里最有价值的部分是坑位清单:每次你不得不纠正 agent,就补一行 你的 agent 刚做完一件你常交给它的任务时,就在同一个会话里把下面这段发给 Claude Code,把你怎么做的存成标准技能 👇 「按照 https://t.co/YN17Wv07wD,把我们刚做的这个任务变成一个技能。只使用这次对话里跑通的步骤、我纠正过你的地方,以及 CLAUDE.md、AGENTS.md 和 git 历史里的相关修正;不要用通用知识填补空缺。把描述写成『use this skill when...』,包含我没有明说但应该触发它的情况,并保持在 1,024 个字符以内。SKILL.md 保持在 500 行以内,细节移到 references/ 并注明每个文件何时该读,给坑位单独开一节。名字用小写字母、数字和连字符,与文件夹名一致。最后,在 evals/evals.json 里写 2-3 个真实测试用例(prompt 加预期输出),每个都跑一遍,把结果和你的操作展示给我。」
我们正在把 Agent Skills 捐给 Agentic AI 基金会(@AgenticAIFdn),也就是我们捐赠 MCP 的同一个基金会。 自 2025 年 12 月以来,已有 70+ 个 agent 平台集成了这一标准,包括 ChatGPT、Cursor、GitHub Copilot 和 VS Code。 这对你意味着什么:
深度解析
是什么 Vox 转述 Anthropic 把 Agent Skills 标准捐给 Linux Foundation 旗下 Agentic AI Foundation(与 MCP 同一归处),并给出官方《Best practices for skill creators》的六条要点,末尾附一段可直接粘进 Claude Code 的提示词,把刚完成的任务转成标准技能。
核心论点链条 1)技能价值来自真实专长,不是模型通识:先在会话里真做一遍任务,再提取成功步骤、被纠正处、输入输出格式与补充上下文;空隙不能用通用知识填。2)每条内容过一遍「没有它 Agent 会做错吗」,否则删掉。3)启动时只加载 name 与 description,所以描述决定技能是否被激活——写成「use this skill when…」,偏激进,列举用户不会明说的触发场景,控制在 1024 字符内。4)SKILL.md 控制在 500 行 / 5000 token 内,细节下沉到 references/ 并注明何时读哪个文件,即渐进式披露。5)具体度对齐脆弱度:脆弱步骤硬编码命令,有判断空间处解释原因,多路径时给一个默认。6)最值钱的是 gotchas 清单:每次纠正 Agent 就记一行。官方文档还补了验证循环、plan-validate-execute、脚本打包等模式。
关键洞察 技能不是文档,而是「上下文投资」:每个 token 都在和对话历史竞争注意力;而 description 相当于技能的召回率参数,写得谦逊就等于没人用。把纠正记录固化成资产,是人与 Agent 协作中最容易被忽略的复利。那份提示词把整套方法论压成一次可执行操作,并要求用 evals 自证效果,是这套规则的实用封装。
边界与风险 一年百万技能、2.8 亿次安装与 70+ 平台支持均为原文数据,未经独立核实;「标准捐赠」对实际互操作性的影响尚待观察;提示词生成的技能仍需人工审阅与多轮迭代,一次执行常常不够。
精选追评与案例
指出官方提示只覆盖调用后的输出、缺少触发层评估,是可落地的边界补强。
@Voxyz_ai those evals check the output once the skill is invoked, but what its missing is the trigger evals. I would add trigger evals per model too.
中文译文 · 机器翻译
这些 evals 检查的是技能被调用之后的输出,但缺的是触发评估。我会再为每个模型各加一套触发评估。
补充实操节奏:趁刚做完把纠正记录下来,比隔几天复盘更能写出有用技能。
@Voxyz_ai 刚做完就把纠正过的地方留下来,最容易写出有用的技能。隔几天再总结,就容易只剩一套漂亮流程。