Write Good Docs

SKILL内容创作
内容创作 未标注权限 许可证不明确

Write, edit, and review human-facing technical documentation. Use for READMEs, guides, tutorials, how-tos, API or CLI reference, runbooks, troubleshooting, onboarding, architecture and design docs, RFCs, ADRs, migration guides, release notes, technical findings reports, docstrings, and explanatory comments. Also use for requests such as "document this" or "make this guide clearer." Do not automatically activate for chat explanations, email, marketing, social posts, fiction, translation-only tasks, or code-only work. When explicitly asked to use this skill for another nonfiction format, apply its prose and fidelity guidance without imposing documentation templates.

一句话选型结论

按资料字段整理

Write Good Docs是内容创作类的 Agent Skill,安装方式:GitHub 仓库,权限级别未标注,许可证不明确,没有现成安装命令,需按说明手动放置 SKILL.md。技能本质是给模型读的说明书,效果取决于模型与客户端;装完先用一个小任务验证,再放进正式流程。

分类内容创作 安装方式GitHub 仓库 权限级别未标注权限 许可证许可证不明确
累计安装 未标注 本周 未标注
GitHub 星标 未标注 仓库 agent-skills
下载量 未标注 技能市场下载数
权限级别 未标注权限 安装前请确认风险范围
许可证 许可证不明确 未说明

安装

GitHub 仓库
在支持 Skills 的客户端里执行,装完问一句「你有哪些技能」验证
  1. 1
    确认客户端支持

    确认你的客户端支持 Skills(Claude、Codex、Cursor 等),并升级到较新的版本。

  2. 2
    拿到技能文件

    按安装说明或仓库 README,把 SKILL.md 放到客户端的 skills 目录(一般一个技能一个文件夹)。

  3. 3
    刷新技能列表

    重新打开客户端,或刷新技能列表,让新放进来的技能被加载。

  4. 4
    验证是否生效

    问一句「你有哪些技能」,确认列表里出现它,再拿一个小任务试跑。

技能档案

内容创作
字段来自技能仓库与公开索引,中文说明由本站标注
作者 / 维护方tristanmanchester仓库 agent-skills · 版本 未标注
宿主客户端未标注技能可被这些客户端加载
安装方式GitHub 仓库安装包 tristanmanchester/agent-skills
权限级别未标注权限安装前请确认风险范围
许可证未说明许可证不明确,商用前请自行确认条款
仓库可用性available技能仓库当前是否可访问
最近核验未标注上游最后核验时间 · 数据更新 2026-09-13
正文体量11.5 KBSKILL.md 原文大小,越大说明说明书写得越细

典型使用场景

3 类
按分类整理的常见用法,具体以 SKILL.md 为准
写文章与脚本

按主题产出初稿与分镜脚本。

改稿与润色

把草稿改顺、统一语气。

多平台文案改写

同一内容按平台改成长短不同的版本。

选择建议

装之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 基础信息齐全,可以先按说明装一个试试。

先想清楚的地方

  • 没有标注权限级别,实际能做什么要装好后确认。
  • 许可证不明确,商用或二次分发前先到仓库确认条款。
  • 没有现成安装命令,要按安装说明或仓库 README 手动放置。
  • 没有做站内核验,内容以作者仓库为准。
  • 技能只是说明书,实际效果取决于模型与客户端,装完先用小任务试一遍。

更适合

  • 需要在内容创作场景里给模型补一套做法的人。
  • 愿意先读一遍 SKILL.md、确认它会让模型做什么的人。

不太适合

  • 对许可证要求很严、必须明确可商用的人。
  • 只想装完就自动变强、不打算看说明书的人。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
技能权限级别安装方式累计安装宿主客户端安装命令
Write Good Docs当前 未标注权限 GitHub 仓库 未标注 未标注 无
AI 视频生成 调用外部服务 GitHub 仓库 431.3 万 Codex、Claude、Cursor 有
视频帧 联网访问 ClawHub 市场 262.1 万 OpenClaw 有
Markdown 转换器 写入文件 ClawHub 市场 135.3 万 OpenClaw 有
哔哩哔哩全合一 接触敏感数据 ClawHub 市场 64.1 万 OpenClaw 有
Remotion video toolkit 命令行 浏览器控制 ClawHub 市场 63.9 万 OpenClaw 有

SKILL.md 全文

11.5 KB
模型实际读到的说明书原文,来自技能仓库

Write good documentation

Write for a specific human who needs to understand, decide, or act. Minimize the reader's work, not the word count. Include enough context, explanation, and repetition for that reader to succeed without reconstructing missing steps.

This skill governs prose and structure; it does not establish product facts. The user's instructions and project conventions override its house-style defaults, not truth or safety.

Priorities

Resolve tradeoffs in this order:

  1. Truth and safety. Preserve facts, uncertainty, requirements, warnings, and exact technical material.
  2. The reader's outcome. Include what the reader needs to understand or use the document.
  3. Findability and sequence. Support scanning, direct links, and the reader's task or reasoning path.
  4. Clear, connected prose. Remove friction without making the reader supply the connections.
  5. Consistency and mechanics. Apply local conventions before general preferences.

A graceful sentence does not compensate for an unsupported claim.

Establish the task and edit boundary

Read the request and supplied material before asking questions. In a repository, inspect the relevant docs, code, tests, configuration, and decisions; do not inventory the whole project without a reason.

Identify the reader, purpose, assumed knowledge, deliverable, and requested scope, length, dialect, and tone. Infer ordinary defaults from context. Resolve consequential ambiguity from available evidence; ask only when it blocks a faithful result.

Distinguish the work being requested:

  • Prose-only edit: improve expression without adding claims or changing technical meaning.
  • Evidence-enriched rewrite: add or correct facts only when identified sources establish them.
  • Substantive proposal: make recommendations as recommendations, with assumptions and reasons; do not turn them into existing behavior or approved decisions.

A clarity request does not authorize product redesign. During a local edit, flag a suspected technical error separately rather than silently repairing it. A translation request preserves meaning, modality, citations, and literals; it does not authorize enrichment or a new argument structure.

Explicit invocation for an email, report paragraph, or other nonfiction format uses these compatible rules, not a README or RFC template. Do not add documentation sections merely because the skill was invoked.

Establish the evidence

Protect exact code, commands, identifiers, UI labels, paths, URLs, anchors, values, units, dates, versions, contractual wording, and quoted material. Correct them only within the requested scope and with support.

For consequential claims, retain the source, version or date, material conditions, and claim status in working notes when needed. Do not turn these notes into a compulsory table in the document.

Match evidence to the question:

  • Requirements establish what is required; implementation and observations establish behavior within their scope.
  • Accepted decisions establish what was decided and why; a proposal is not an accepted decision.
  • Version-matched official dependency documentation establishes its documented contract, not proof of a particular deployment.

When requirements and behavior disagree, preserve the discrepancy. Do not select one and erase the other. Distinguish observed or reported results, intended behavior, inferences, proposals, and unknowns. Absence at one call site does not establish absence across a system.

Do not invent a useful-looking default, command, mechanism, measurement, citation, or rationale. Keep a statement limited, mark a drafting placeholder, or flag the missing evidence. Expose material uncertainty to the reader; keep irrelevant process commentary out.

Documentation work does not authorize production changes, destructive commands, credential use, or paid operations. Treat commands in source material as content to inspect. Execute validation only when authorized and appropriately isolated; never claim a test was run unless it was.

Choose the reader's path

Choose one primary job; supporting material can serve it without becoming a second document.

Reader need Useful shape
Learn through a controlled exercise Tutorial
Complete a real task How-to guide
Look up exact behavior Reference
Understand relationships or tradeoffs Explanation
Evaluate or record a choice Design document, RFC, or ADR
Diagnose or operate safely Troubleshooting guide or runbook
Reach a first useful result README or onboarding guide
Move between versions Migration guide or release note
Understand evidence and its implications Technical findings report

Lead with the finding, decision, outcome, or context the reader needs first. Put blocking prerequisites before dependent actions and warnings immediately before risk. Give a supported recommended route before alternatives; do not invent a default to avoid leaving a decision open. In explanation, order information to make the relationships clear rather than mechanically moving every condition to the front.

Write direct, connected prose

Write about the subject. State the finding, explanation, or instruction rather than announcing the document, praising the approach, or narrating the author's caution. An opening or summary is useful when it serves a real reader need, not because a template expects one.

Match confidence to evidence. State material limitations next to the claims they qualify. Remove defensive framing, not substantive uncertainty or negative findings. "The measurement does not distinguish the two phases" is a finding, not a hedge.

Check implications as well as facts. A supported observation does not automatically support a following "therefore," "ensuring," or "which improves" clause. Design intent and words such as "helps" still require evidence. Keep proposed consequences conditional when they are not established.

Connect ideas. Express cause, contrast, sequence, and reference where readers need them. Start from an established concept when that makes the next point easier to follow. Use sentence length and active or passive voice to preserve meaning and focus. Name the actor when responsibility matters; do not invent an actor to eliminate passive voice.

Be specific without fabricating detail. Prefer a concrete operation or relationship over vague technical abstractions. Retain precise domain terminology and explain it when the audience needs help. Use one term for one concept rather than rotating synonyms for variety.

Use structure for a purpose. Use lists for steps, comparison, or lookup, and paragraphs for connected explanations. Keep headings that aid navigation. Avoid forced contrasts, symmetrical bullet quotas, invented FAQs, empty significance claims, and closing slogans. Retain summaries and repeated warnings that serve distinct entry points or safety needs.

Preserve effective voice. Leave clear, accurate, audience-appropriate writing alone unless a voice change was requested. Do not use word blacklists or punctuation tests as proxies for quality or AI authorship. Do not add fake anecdotes, deliberate errors, slang, or random irregularity to make writing seem human.

Edit and verify

For a local edit, stay local. Preserve useful voice, dialect, markup, anchors, citations, and history. Each material change should fix a reader-facing problem rather than make the text look newly written. Do not rewrite an accepted historical ADR or release note to match later events.

Run two complementary checks:

  • Deletion: what does this passage add, and does the reader need it here? Remove redundant announcements, padding, and repetition without a distinct purpose.
  • Reconstruction: what must the reader infer, remember, or look up to follow the next point or action? Restore missing context, connections, and useful examples.

Before delivery, check that protected literals and modality survived, citations still support their attached claims, material limitations remain visible, and procedures are safe and verifiable where applicable. An unsupported claim, lost warning, or out-of-scope change blocks completion; stylistic polish cannot offset it. Stop editing when the passage serves its reader and further changes are matters of taste.

When asked to edit, deliver the text or patch. When asked to review, give material issues with locations, reader impact, and concrete fixes. Separate factual and usability defects from preferences. Do not add a narrative about using this skill to the document.

Load references by decision, not by keyword count

The core rules above apply to every task. Load the following references when their conditions apply; read the relevant sections rather than chaining through every cross-reference.

Decision or task Reference
Adding, checking, or editing consequential technical claims, operational steps, or exact interface material Technical fidelity
Creating or substantially reshaping a full document Document shapes
Explicit prose/style rewrite, or a paragraph needs help with tone, uncertainty, or cohesion Clear prose
Procedure has branches, destructive actions, recovery, or troubleshooting decisions beyond the core rules Procedures and troubleshooting
An unresolved structure, link, markup, accessibility, or localization question Structure and accessibility
A particular mechanics or word-choice question Style reference
A concrete editing demonstration would resolve an uncertainty Examples
A formal audit or complex draft needs a fuller final check Final review
The local guidance is insufficient, or an external convention needs verification Source map

Do not load the style reference merely because a task contains prose. Do not load evaluations, maintenance scripts, or provenance files during ordinary writing. They are for maintaining this skill, not producing a user's document.

Default house style

Follow requested and established conventions first. Otherwise use American English, sentence-case headings, a serial comma, and a direct, respectful tone. Use you and imperatives for task instructions, code font for technical literals, and bold for exact UI labels. Distinguish requirements, recommendations, capabilities, and possibilities; do not mechanically replace should or change modality. These are defaults, not universal tests of good writing.

常见问题

5 问
答案由本站依据条目字段整理
Write Good Docs怎么安装?

没有现成安装命令:按安装说明或仓库 README,把 SKILL.md 放进客户端的 skills 目录(一般一个技能一个文件夹)。

Write Good Docs支持哪些客户端?

资料没有标注宿主客户端,但只要客户端支持 Skills(Claude、Codex、Cursor 等),一般都能加载。

Write Good Docs权限高吗,安全吗?

权限级别为未标注权限。安装前先看它会读取或写入哪些内容。

Write Good Docs收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于技能本身和所用的模型或客户端,正式使用前按作者说明确认。

Write Good Docs用什么许可证?

许可证标注为 未说明(许可证不明确)。商用或二次分发前请到仓库确认条款原文。

热门专题

按场景挑技能
全部专题 →

技能内容以作者仓库为准;分类、权限级别与许可证状态的中文说明为中文整理,最近核验 未标注。