1实现原理 · 为什么它能做到
纯「参考文档」型 skill,无任何可执行脚本/工具调用/网络请求——功能全部靠「被 agent 读进上下文后改变其写作行为」实现,对写作者(人或 agent)起尺子作用,不起执行器作用。
Reference for writing any document an agent consumes: a skill, an `AGENTS.md` / `CLAUDE.md`, a doc reached by a pointer. The packaging differs; the writing does not: the same levers make each one predictable, since the agent takes the same _process_ every run rather than producing the same output.
触发机制是模型自动调用(model-invoked):frontmatter 保留 description 且未设 disable-model-invocation,description 即技能最顶层的 context pointer,其措辞决定 agent 何时取用该材料。
description: Writing documents for agents. Use when creating or editing skills, or modifying AGENTS.md or CLAUDE.md.
采用渐进披露结构:技能专属内容(frontmatter、invocation 选择、router skills)拆到 SKILL-MECHANICS.md,仅当要写的是 skill 时才经相对链接取用,压低主文件的常驻负载。
When the document you're writing is a skill, read [`SKILL-MECHANICS.md`](SKILL-MECHANICS.md) for frontmatter, invocation choice, and router skills.
核心行为锚点是『leading word』:刻意选用模型预训练中已存在的概念词(_lesson_、_fog of war_、_tracer bullets_)作为单 token 行为锚,重复 token 而不重复句子,用最少 token 调动模型已有先验来固定一片行为区域。
A **leading word** is a compact concept already living in the model's pretraining that the agent thinks with while running the document (_lesson_, _fog of war_, _tracer bullets_). Repeated as a token, never as a sentence, it accumulates a distributed definition and anchors a whole region of behaviour in the fewest tokens, by recruiting priors the model already holds.
剪枝判据是行为主义的 no-op 测试而非篇幅美学:一句指令若模型默认就遵守,就是付了 load 却啥也没说的 no-op,删除时应整句删除;两句分歧以『跑一遍文档看行为变没变』裁决,不以辩论裁决。
Hunt **no-ops** sentence by sentence: an instruction the model already obeys by default pays load to say nothing. The test (does it change behaviour versus the default?) is model-relative, not reader-relative: two people disagreeing about a no-op disagree about the default, and settle it by running the document, not by debate. When a sentence fails, delete the whole sentence rather than trim words from it.
用『两负载』(context load 与 cognitive load)作为统一决策经济学,把拆分/内联/披露/pointer 措辞等所有写作决策化约为同一个权衡的不同位置。
**Context load** is the cost of always-loaded material on the agent's window: an `AGENTS.md` line, a skill description, anything sitting in context every turn, spending tokens and attention whether or not it fires. / **Cognitive load** is the cost on the human: which documents exist and when to reach for each. The human is the index.
2核心能力
4风险提醒 风险提醒:绿色 · 放心使用
- 纯咨询、无校验:效果依赖执行者判断力,新手可能机械执行『删句/去重』而误删合法内容 — no-op 判据是行为测试('settle it by running the document'),文档作者通常没有条件真正跑一遍,容易退化成凭感觉删减
- 抽象术语密集、思维模型多(pointer/两负载/信息层级/criteria/leading word/pruning),上手门槛高 — docs 页专门用一整节回答『English isn't my first language. Do I lose the leading-word advantage?』,侧面印证该担心是真实用户反馈
- description 触发分支窄,模型在写 README/spec/ticket/运行时 prompt 等场景不会自动唤起,容易漏用 — description 只写 'creating or editing skills, or modifying AGENTS.md or CLAUDE.md';docs 页建议其余场景手动输入技能名
- 对『被审阅文档含恶意/对抗性指令』无任何防护指引 — skill 不指示执行文档内容,但若 agent 在读取不可信文档时套用其建议,理论上有被文档内文本操纵的可能(同类文本型 skill 通病)
5第二遍独立确认
- [ok] 外部资源/工具 = 无(技能文件内零 CLI/API/网络端点) — 对 SKILL.md、SKILL-MECHANICS.md、agents/openai.yaml 全文 grep 'http://|https://|curl|subprocess|exec\(|\$\(' 无命中(exit=1)。docs/productivity/writing-for-agents.md 含 aihero.dev 外链,但该页属仓库人类阅读的文档树(AGENTS.md L17:docs 树镜像、URL https://aihero.dev/skills-<skill-name>),不随技能安装、不被 agent 作为技能内容读取,不计入技能外部依赖。
- [ok] 安全结论 green:无脚本、无网络、无凭证、无写文件 — git ls-tree -r HEAD skills/productivity/writing-for-agents/ 恰好 3 个文件(SKILL.md、SKILL-MECHANICS.md、agents/openai.yaml),无隐藏文件/脚本/二进制/符号链接;openai.yaml 仅 3 行 interface 文本。反例搜索无发现。
- [ok] model-invoked 声称(description 在、无 disable-model-invocation) — SKILL.md frontmatter 复核:有 description 行、无 disable-model-invocation 键;与 SKILL-MECHANICS.md L9 给出的 model-invoked 机制定义(omit disable-model-invocation + model-facing description)完全一致。
- [ok] SKILL-MECHANICS.md 经相对 pointer 披露且文件真实存在 — SKILL.md L8 相对链接 SKILL-MECHANICS.md,目标文件在该 pin 树内存在(ls-tree 证实),SKILL-MECHANICS.md 首句反向链回 SKILL.md,互指闭合。
- [ok] 『leading word 调动预训练先验』是否夸大 — 复核原文:主张以『compact concept already living in the model's pretraining』锚定行为,属行文设计主张而非可执行承诺;全文件没有任何自动化/eval/可测产出声称(docs 页反而自述 'There is no automated eval here'),无吹牛。
- [ok] description 范围(skill/AGENTS.md/CLAUDE.md)与正文范围(一切 agent 文档)的出入 — 非缺陷:按该 skill 自己的 pointer 经济学(常驻 description 每个词都花 context load、pointer 措辞决定触发),收窄分支是刻意设计;docs 页明示其余场景 'Reach for it by hand',人可随时输入技能名。
- [ok] 更名史(writing-great-skills → writing-for-agents, v1.1) — docs/productivity/writing-for-agents.md(同 pin 仓库内):'It was called `writing-great-skills` until v1.1.' 及 'There is no alias. Reinstall under the new name.';技能目录内 SKILL.md 未提别名,与『无 alias』一致。克隆为浅克隆(depth=1),无法用 git log 交叉验证 rename commit。
- [ok] stars/许可证/推送时间元数据 — GitHub API 两次拉取一致:stargazers_count 253950、license MIT(仓库 LICENSE 文件头 'MIT License / Copyright (c) 2026 Matt Pocock')、pushed_at 2026-09-04T08:45:43Z(与 pin commit 时间 2026-09-04 09:43:27+0100 同一天)。注意:仓库创建于 2026-02-03,星标为 2026-09-06 快照,仅对该时点有效。
6结论
2741c227a60b8035…3cca18b368