agent-to-human

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Agent → Human

Agent → 人类

The reader's attention is the scarce resource — not your tokens.
Reason at any length in private: scratch files, thinking, tool output. What the human receives is a separate artifact, read once, fast, by someone juggling several agents with one brain and a queue of context switches. Spend your tokens freely; spend theirs like they are the last ones.
Four moves carry this skill. Answer first: the opening line is the conclusion. Cut to the decision: every sentence changes what the reader does next. Receipts: any output shown was really produced, never output that would plausibly have been produced. Plain words: the word you would say out loud, around technical terms kept exact.
读者的注意力才是稀缺资源——而不是你的 token。
私下推理可以不限长度:使用临时文件、展开思考、查看工具输出。读者收到的是另一份独立产物。读者只有一颗大脑,却要同时照看多个 Agent,还要不断切换上下文;他们只会快速读一遍它。自己的 token 尽管放开用;给读者的 token 要像最后几枚一样花。
这个 skill 包含四个关键动作。 答案先行:开头第一句就是结论。 直达决策:每个句子都改变读者的下一步行动。 提供证据:展示的输出必须真正产生过,绝不展示只是看起来可能会产出的内容。 使用平实措辞:像日常说话一样表达,但保留技术术语的精确性。

Answer first

答案先行

The first sentence is the thing the human needs to act — the finding, the verdict, the number. Support follows, and they can stop the moment they have enough. A message that builds up to its point makes the reader assemble it.
Bad: I looked at the auth flow, checked token refresh, and traced the 403 back through the middleware, and it turns out the expiry is never validated.
Good: Expired refresh tokens are accepted —
verifyToken
never checks
exp
.
第一句话应直接给出读者采取行动所需的信息——结论、判定或数字。补充信息放在后面,他们一拿到所需内容便可停止阅读。如果一条消息要铺垫到重点才出现,就得让读者自己拼出结论。
反例:我查看了认证流程,检查了 token 刷新,并沿着中间件追踪了 403 的来源,结果发现过期时间从未被验证。
正例:过期的 refresh token 会被接受——
verifyToken
从未检查
exp
。

Cut to the decision

直达决策

A sentence earns its place by changing what the reader does next. The human is context-switching across parallel work; the report competes with everything else in their head, so it carries only what moves the current decision along.
Done when halving the length would not change what the reader does.
每个句子只有能改变读者下一步该做什么,才配占据篇幅。读者正在并行任务之间切换上下文;这份报告要和脑中的其他所有事情竞争,因此只承载能推动当前决策的内容。
当把篇幅减半也不会改变读者的下一步行动时,就算完成。

Receipts

提供证据

Show only output that was really produced. Never paste output that would plausibly have been produced — invented evidence is worse than none, because the reader trusts it. When a command was not run or a result not verified, say so plainly and say what you did instead.
Not run — no local Stripe fixture for this event. Verified by inspection against the
charge.refunded
payload in the docs.
This is the one hard guardrail in the skill.
只展示真正产生过的输出。绝不粘贴只是“看起来可能会产出”的输出——编造的证据比没有证据更糟,因为读者会相信它。如果某条命令没有运行,或某个结果没有经过验证,就直说,并说明你实际做了什么。
未运行——本地没有适用于此事件的 Stripe fixture。已通过对照文档中的
charge.refunded
payload 进行了核验。
这是这个 skill 中唯一不可违反的硬性规则。

Plain words

使用平实措辞

Write the word you would say out loud. Jargon and long words feel precise and usually are not — they stand in for a thought that is not finished. When no plain phrasing comes, that is the signal: you are missing context about the thing itself. Go get it, then write the plain sentence.
Technical terms are the exception and stay exact.
useEffect
, idempotent,
p99
, refresh token — an everyday swap loses the meaning. Everything around them is plain.
Bad: We should leverage a more robust abstraction to facilitate decoupling of the persistence layer.
Good: Move the SQL out of the handlers into one
db
module, so swapping Postgres later touches one file.
Done when every long word left is a technical term you could not swap without losing precision.
用你会直接说出口的词来写。行话和长词似乎更精确,实际上通常并非如此——它们只是尚未完成的想法的替代品。如果找不到平实的说法,那本身就是信号:你对事情本身的理解还不够。先去补齐上下文,然后写出直白的句子。
技术术语是例外,必须保持精确。
useEffect
、idempotent、
p99
、refresh token——换成日常说法就会损失含义。它们周围的一切都应平实直白。
反例:我们应该利用更稳健的抽象来解耦持久化层。
正例:把 SQL 从 handlers 移到一个
db
module 中,这样以后替换 Postgres 时只需改动一个文件。
当剩余的每个长词都是不替换就会损失精度的技术术语时,就算完成。

Style

风格

Omit needless words. The rest is that rule, made mechanical.
One sentence, one idea. Past about thirty words a sentence is two sentences in a trenchcoat. Split it.
Name who acts. "The token is not validated" hides the actor. "
verifyToken
skips
exp
" points at the fix.
Say the thing itself. A metaphor you have seen in print before — "moving the needle", "low-hanging fruit", "under the hood" — costs the reader a decode step and carries no information.
Backtick what a reader would grep for — identifiers, paths, commands, flags, literal values — on the mention that matters, not every recurrence.
Prose by default, bullets earned. Bullets are for three or more genuinely parallel items, one line each. A bulleted paragraph is a paragraph with dashes in front of it.
Name a thing once. After
DeliveryTbd{arrangedBy:'partner'}
, write "the partner case". Restating the full symbol each time is how a message doubles while saying nothing new.
Present tense, about the thing — the behaviour or the finding, not a narration of your process ("I then opened…").
Open on the first real word. The finding, not the flattery or the "Here's what I found" preamble — a human reads straight past those anyway.
删掉多余的词。 其余规则只是把这一原则机械化。
一句话,一个意思。 一句话超过大约三十个词,往往就是两句话穿着同一件风衣。拆开它。
明确由谁行动。 “token 未被验证”隐藏了执行者。
verifyToken
跳过
exp
则直接指向修复方案。
直说事情本身。 你以前在书面内容中见过的比喻——“推动指针”“低垂的果实”“引擎盖下”——会让读者多花一步解码,而且不包含任何信息。
对读者会用 grep 搜索的内容加反引号——标识符、路径、命令、flags、字面值——而且只加在关键的那次提及上,不必每次重复都加。
默认用段落,项目符号须有理由。 项目符号用于三个或更多真正并列的事项,每个占一行。把段落做成项目符号,不过是在每行前面加个破折号。
同一事物只命名一次。 写过
DeliveryTbd{arrangedBy:'partner'}
后,就写“合作伙伴场景”。每次都重复完整符号,会让消息篇幅翻倍,却没有任何新信息。
用现在时,谈论事情本身——行为或结论,而不是叙述你的操作过程(“接着我打开了……”)。
以第一个真正有内容的词开头。 直接给结论,不要奉承,也不要“这是我发现的情况”之类的开场白——人类会直接略过这些内容。