alber70g/skills

writing-clearly

Keeps any text written for someone else plain, concrete and verified, so the reader gets the point at first read: replies to the user, documentation, slides, website and UI copy, emails, messages, commit messages, spoken notifications.

查看源码

使用方法

  1. 复制安装命令,在自己的智能体环境安装前先检查源码包。
  2. 从原始文档核对所需工具、凭据和支持的宿主。
  3. 说明输入材料与期望交付物,并对照来源检查输出。
npx skills add https://github.com/alber70g/skills --skill writing-clearly

开始前需了解

  • 安装不包含模型订阅、API 额度或访问文件的权限。
  • 仓库星数和安装量反映活跃情况,不等于独立质量或安全测试。
仓库原始内容

按源仓库内容呈现,保留标题、案例、代码、表格、链接以及原文引用的演示图片。

Writing Clearly

For anyone who writes or edits text with or for the user: documentation, slides, web and UI copy, emails, messages and commit messages. Collected from writing instructions and from corrections made during a review of training material.

The reader should get the point at first read, trust that it's accurate, and know what to do with it. A vague phrase feels finished to the writer, who already knows what it means; the reader has to find out what it means.

Contents

Rules

  1. Say what it is. Titles and labels describe their content, like a documentation heading. No

slogans, puns or statements the reader has to decode.

  1. Describe, don't sell or nag. Everywhere, not only in titles: say what something is or does,

not that it's good, important or a must. Test: negate the qualifier. "Writing skills that work" becomes "Writing skills that don't work"; nobody writes that, so the qualifier adds nothing: "Writing skills". If the negated version is something people actually do or believe ("review only the code"), the qualifier draws a real distinction and stays.

  1. Name the topic it belongs to, when a piece is part of a group: "Skills: writing

descriptions". Leave the prefix out when it adds nothing, and don't repeat the topic word after it.

  1. The title doesn't repeat the body. If the list of questions is on the slide, the title is

"Introduction", not "Introductions: name, role and three questions".

  1. Reason first, in one line. Then the guidance.
  2. Concrete over figurative. The exact command, number, name, place or step, not a turn of

phrase that stands in for it.

  1. Professional register. This is not a poem, a rap or a chat with classmates. Use established

terms with no hidden meaning: no wordplay, metaphors, dramatic lines, punchlines or in-group slang. Established technical terms are fine ("lethal trifecta", "red/green", "slopsquatting").

  1. Only state what's checked.
  • Verify facts against a primary source before they go in; mark anything unverified as such.
  • No rankings or superlatives nobody measured ("the most important", "the highest-value").
  • Read a file, slide or page before describing it; otherwise say "not checked".
  • Don't assume people's roles or pronouns. Use "they" unless pronouns are stated.
  1. What the reader acts on is visible, not tucked in notes, links or footnotes. Background can

go in speaker notes, linked pages or an appendix.

  1. Every line adds something new. No filler, no repetition, no hype.
  2. Same patterns throughout: naming, structure, emphasis.
  3. No italics in titles. They're obtrusive. Use bold in body text for the word that carries

the point.

  1. When editing someone else's material, move rather than delete. Content that doesn't fit

one place (speaker notes, background) goes somewhere else, not away. Keep facts, names, numbers and quotes intact.

  1. Last check before sending: could the reader act on each sentence without asking what was

meant? If not, replace the phrase with the plain fact it stands for.

Messages and emails

When the user asks to make a message sendable, distill a text or write an email to a group, read references/messages-and-emails.md first.

Before and after

Corrections from real use, grouped by the rule they fall under. Add new ones as they're caught.

Say what it is (titles and labels)

BeforeAfter
"A skill that triggers" / "Will it fire?""Skills: writing descriptions"
"Gate 1: ask for the evidence, not the claim""Gate 1: ask for the test output"
"Don't reinvent the wheel""Libraries: adopt, extend, fork or avoid"
"The day's spine (know it cold)""Structure of the day"

Describe, don't sell or nag

BeforeAfter
"Writing skills that work""Writing skills"
"Context discipline""Context management"
"Never accept a bare 'done'.""Don't accept 'done' without that evidence."
"Memorize these, because the exercise grades on them.""Four properties; the exercise grades on them."

Prefixes and titles that repeat the body

BeforeAfter
"Introductions: name, role and three questions" (the questions are on the slide)"Introduction"
"Instruction layers: …" (prefix adds nothing)"CLAUDE.md instruction layers"
"Test a skill like code" (in a group of skill slides)"Skills: test them like code"

Only state what's checked

BeforeAfter
"The highest-value line you can write." (unmeasured, no reason, no place)"Add this to CLAUDE.md: an edit that looks isolated can break callers the agent never read."
"The most important thing we saw was…""The topic that came up most was…"
"~/.ssh and ~/.aws unreadable inside the sandbox" (checked: wrong)"~/.ssh and ~/.aws are readable by default until you add credential entries."
"He" for someone whose pronouns are unknown"They"
Describing a slide without reading itRead it first, or say "not checked"

Concrete over figurative

BeforeAfter
"Before it goes out""After writing, before it's sent, published or committed"
"the agent walks in blind""the agent starts without instructions"
"understand the blast radius first""check what depends on the code before changing it"
"A criterion without a command is a wish.""A criterion without a command can't be verified."
"Background can live a layer down""Background can go in speaker notes, linked pages or an appendix"
"Every line does work" / "dresses things up""Each line adds something new" / "no hype"

Professional register (slogans, punchlines, drama)

BeforeAfter
"Generation got cheap. The verdict is your job.""Agents write code quickly. Deciding whether it is correct is your job."
"A claim is not an artifact.""The agent's summary is not evidence. Read the tool output."
"CLAUDE.md asks. A hook runs every time.""CLAUDE.md is an instruction the agent can skip. A hook runs on every stop."
"Still green? Theater.""Still green? The test can't detect this defect."

Feedback loop

When the user points out vague or guessed wording:

  1. Rewrite the phrase plainly, right away, in the text it came from.
  2. Add the pair to the "Before and after" table where the rules live for this user: their

CLAUDE.md / AGENTS.md if the rules were copied there, otherwise this file.

  1. If the agent has persistent memory, save or update one feedback memory with the phrase, the

plain version and why it was vague. Update the existing memory rather than adding a new one per slip.

  1. If the same kind of slip recurs, sharpen the rule it falls under so the rule covers it.

按任务相关推荐

来自同一仓库

更多 Skills

全部 Skills
alber70g
社区

general-guidelines

Behavioral guidelines to reduce common LLM coding mistakes. Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes, surface assumptions, define verifiable success criteria, keep a WORKLOG.md so work can be continued across sessions, and write research to research/ and plans for review to docs/.

1小时安装量
2
GitHub Stars
0
最近更新
10月18日
alber70g
社区

orchestrator

Session mode where the main thread orchestrates: subagents on a cheaper model do research and execution, the main thread plans and validates. Use when the user runs /orchestrator, asks you to act as an orchestrator, asks to keep the main thread token-efficient, requests a large implementation plan or a large research task, or wants you to work toward an overarching goal.

1小时安装量
1
GitHub Stars
0
最近更新
10月18日
alber70g
社区

turn-based-discovery

Use the AskUserQuestion tool (the same multiple-choice mechanism plan mode uses) to get clarity from the user. Each question is multiple-choice plus a free-format "Other" escape. Ask in small batches — up to 4 questions per call (the tool's max), more if needed across calls. Use when scoping work, resolving ambiguity, or surfacing answers a research pass already raised. Triggers: /turn-based, ask me, clarify, discovery mode, interview me.

1小时安装量
2
GitHub Stars
0
最近更新
10月18日
alber70g
社区

autoresearch

Autonomous experimentation skill — agent interviews the user, sets up a lab, then explores freely (think, test, reflect) until stopped or a target is hit. Works for any domain where you can measure or evaluate a result.

1小时安装量
1
GitHub Stars
0
最近更新
10月18日