跳到正文
cccskills
免费在 GitHub 公开

great-writer

Universal writing skill for AI agents. 9 writing modes (tech articles, marketing copy, research reports, Xiaohongshu notes, technical docs, GitHub READMEs, creative writing, rewrite/polish, editorial review) + 9 core modules (research, core-finding, writing DNA, humanizer/polish, review, style learning, adaptation, SEO/GEO, writing memory, visualization). 6-phase pipeline with deep-thinking methodology. Bilingual (EN/ZH).

中文简介正在准备中。当前显示原文说明。

d-wwei/great-writer星标 24(GitHub 24 + cccskills 0)
集成工具GitHub

实测评分 未实测

查看安装方法

包含的文件(28)

  • SKILL.md15.4 KB
  • .gitignore10 B
  • core/adapt-protocol.md4.8 KB
  • core/core-finding.md4.7 KB
  • core/humanizer.md9.7 KB
  • core/research-protocol.md5.9 KB
  • core/review-protocol.md8.3 KB
  • core/seo-layer.md5.7 KB
  • core/style-learner.md4.5 KB
  • core/visualization.md5.1 KB
  • core/writing-dna.md19.8 KB
  • core/writing-memory.md5.0 KB
  • docs/superpowers/plans/2026-03-28-great-writer-v2.md44.0 KB
  • docs/superpowers/specs/2026-03-28-great-writer-v2-design.md18.9 KB
  • great-writer-bundled.md171.6 KB
  • LICENSE1.0 KB
  • modes/creative-writing.md7.2 KB
  • modes/editorial-review.md5.7 KB
  • modes/github-readme.md27.2 KB
  • modes/marketing-copy.md5.1 KB
  • modes/research-report.md6.0 KB
  • modes/rewrite.md5.3 KB
  • modes/tech-article.md18.1 KB
  • modes/technical-docs.md7.4 KB
  • modes/xiaohongshu.md6.8 KB
  • README.md12.0 KB
  • README.zh.md10.4 KB
  • scripts/bundle.sh1.4 KB

SKILL.md(原文)

安装前可以查看交给 AI 代理的指令内容。

Great Writer — Universal Writing Skill

Stance

Thinks from the reader's position, not the writer's. Every sentence earns its place by creating value for the reader — not by documenting what the writer knows. Treats rhythm, asymmetry, and surprise as engineering parameters, not aesthetic preferences. Writes like a human: irregular, opinionated, specific.

心里放一个具体的人,写给他,不是写给"读者们"。先亮自己的弯路,再给方向——所有说服力的来源不是因为你对,而是因为你先错过。

This skill covers 9 writing modes through a modular 6-phase pipeline, with 9 core capability modules. Each piece goes through core logic (research + find core), structure design, drafting, review, polishing, and finalization before output.


Red Lines

Scan all output against these constraints before delivery. Every item is mechanically checkable.

  1. No output without Research Summary — Phase 1 Step a gate. If skipped, mark output [⚠️ Written without research]
  2. No unexamined core — Phase 1 Step c gate. The core must survive an attack before expansion. An unexamined opinion, no matter how well-written, collapses at the first smart reader's question.
  3. No draft without structural outline — Phase 2 gate. Skeleton must exist before prose
  4. No 3+ consecutive sentences of similar length — Adjacent sentences within ±20% word count = low burstiness. Insert a short punch or a long breather
  5. No symmetrical section structure — If all sections have the same paragraph count and internal pattern, restructure until emphasis is visible
  6. No AI-slop phrases — Blacklist in core/humanizer.md Level 1 applies to all output, no exceptions
  7. No unsupported claims — Every claim needs a data point, example, or source. "This is powerful" without evidence = delete
  8. No significance inflation — If a paragraph ends by explaining why the thing matters when the reader can already see it, delete that sentence
  9. No swap-test failure — If you can substitute a different subject and the piece still reads the same, it is too generic. Rewrite with specifics that only apply to this subject
  10. No translation voice — Chinese writing only. Test: translate the sentence back to English, then back to Chinese. Still the same? → Translation voice. Rewrite with Chinese rhythm. (See writing-dna.md Translation Immunity for symptoms checklist)

Acceptance Criteria

From the reader's perspective — every item testable by two independent reviewers.

  1. Actionable endpoint: Reader can identify what to do next (CTA, changed understanding, or specific next step)
  2. Evidence density: At least one data point, analogy, or concrete example per major section
  3. Burstiness: Sentence length standard deviation ≥ 30% of mean sentence length across the piece
  4. Research-dependent: The piece could NOT have been written without the specific research done in Phase 1 — remove all research findings and the piece collapses
  5. Tension: The piece changes the reader's mind about at least one thing, or teaches something the reader did not know and would find surprising

Step 1: Route to the Right Mode

Identify the writing type from the user's request and load the corresponding mode template.

SignalModeFile
公众号, 博客, 技术博客, 产品介绍, 产品发布, 推广文, blog post, product article, tech articleTech Product Articlemodes/tech-article.md
Landing page, 广告, CTA, 社交媒体, 文案, ad copy, social post, marketing copyMarketing Copymodes/marketing-copy.md
白皮书, 行业分析, 竞品报告, 投研, whitepaper, research report, competitive analysisResearch Reportmodes/research-report.md
小红书, 种草, 笔记, xiaohongshu, RED noteXiaohongshu Notemodes/xiaohongshu.md
GitHub README, repo README, project README, 写 README, 仓库介绍, 仓库 README, write a README forGitHub READMEmodes/github-readme.md
API docs, changelog, 技术文档, 内部文档, technical docs, internal guide, release notesTechnical Documentationmodes/technical-docs.md
改写, 润色, 帮我改, polish, rewrite, improve this, edit this, make this betterRewrite / Polishmodes/rewrite.md
审稿, 帮我看看, 编辑审核, editorial review, critique this, review this articleEditorial Reviewmodes/editorial-review.md
写小说, 写故事, 写散文, 写演讲稿, 写剧本, creative writing, fiction, essay, short story, speechCreative Writingmodes/creative-writing.md
Ambiguous or unclearAsk the user to clarify—

After routing: Read the matched mode file (relative to this skill's root directory) to get the type-specific structure template, tone rules, and self-check checklist.


Step 2: Run the Writing Pipeline

Every piece of writing goes through 6 phases. Each phase loads specific modules.

Phase status output (mandatory): When entering each phase, output a brief status line so the user can follow the process. Format:

Phase N: [Name] — [one sentence: what this phase does and why]

Example status lines:

  • Phase 1: Core Logic — Researching the audience, finding the core idea, and stress-testing it before writing a single word.
  • Phase 2: Structure Design — Building the skeleton: a load-bearing analogy, narrative arc, and deliberate weight distribution.
  • Phase 3: Draft — Writing the first draft with 15 principles active and mode-specific voice loaded.
  • Phase 4: Review — Four-dimension check: structure, logic, facts, perspective. Fixing all must-fix items.
  • Phase 5: Polish — Four passes: oral test → density → AI trace removal → anti-style check.
  • Phase 6: Finalize — Running the mode-specific self-check. All items pass → output.

Keep it to one line. The user should know what's happening and why, not the full protocol details. Output in the user's language — if the user writes in Chinese, the status line is in Chinese; if in English, in English.

Phase 1: Core Logic

Three steps to establish what this piece is really about.

Step a — Research. Read core/research-protocol.md and follow its protocol completely.

  • Audience targeting, exclusion analysis, deep material mining
  • Output: Research Summary with killer data points

Step b — Think & Judge. From the Research Summary, identify:

  1. Core tension: What does this piece change about the reader's current understanding? If nothing — the piece has no reason to exist. Reframe until tension is found.
  2. Value carrier: What single element would make readers share this? (A mapping table, a data comparison, an analogy, a causal chain, a key insight.) This element is protected — never compressed to meet length targets.

Step c — Find Core & Attack Core. Read core/core-finding.md and execute the protocol.

  • Use the four shovels on the results from Step a and Step b to dig to the load-bearing idea underneath
  • Attack the core: "If this is true, then why ______?"
  • Three outcomes: core holds → proceed with confidence; core morphs → return to Step a with new angle; core breaks → tell the user honestly
  • Output: Core Statement (one sentence) + Stress-Test Result

Gate: Do not proceed to Phase 2 until Research Summary + Core Statement + Stress-Test Result are all produced.

Phase 2: Structure Design

Design the piece's skeleton around the validated core.

Step a — Scaffold. Find a cross-disciplinary load-bearing analogy for the core.

  • Not decoration — a structural wall. If you remove it, the piece collapses.
  • Search broadly across disciplines: biology, economics, physics, architecture, cooking, sports, games — prioritize domains familiar to the target reader, not just computer science.
  • Validation: (1) Remove it and the piece collapses (load-bearing). (2) Dig one level deeper and it still maps (multi-layer). (3) The reader gets it without explanation (self-evident).

Step b — Narrative arc & weight. Choose the shape and assign weight deliberately.

  • Arc options: problem→solution→evidence, story→principle→application, counterintuitive claim→proof→implications, chronological with twist.
  • Assign weight: which section deserves 40%? Which deserves 5%? Asymmetry is a feature.

Step c — Reader perspective check. Step back and review the skeleton from the reader's viewpoint.

  • Does the structure flow logically for someone who doesn't have your context?
  • Can a reader who only skims headings still grasp the main argument?
  • Present skeleton to user (unless user said "just write it" or the piece is short-form).

Gate: Do not proceed to Phase 3 until the structure skeleton is confirmed.

Phase 3: Draft

Read core/writing-dna.md for universal writing principles (including Principles 10-15 and Density Control). Read the matched modes/{type}.md for the structure template and DNA & Voice Settings.

  • Apply the writing principles from writing-dna.md (check mode differentiation table for active principles)
  • Follow the mode's structure template section by section
  • Hang every detail back on the scaffold — the reader should always know "where this piece fits"
  • Use research findings as the content foundation
  • Output: Complete first draft

Phase 4: Review

Read core/review-protocol.md and run all four review dimensions.

  • Structural review (against mode template)
  • Logic review (argument flow, data support)
  • Fact check (verify numbers, citations, dates)
  • Perspective audit (developer vs product vs user perspective)
  • Output: Numbered fix list → apply all fixes

Gate: Do not proceed to Phase 5 until all "Must Fix" items are resolved.

Phase 5: Polish

Read core/humanizer.md and run all four passes sequentially.

  • Pass 1: Oral test — "Would you say this to a smart friend? No → rewrite." Highest priority. All subsequent passes must re-pass this test.
  • Pass 2: Density / rhythm / word choice / anti-template — apply the five-dimensional compression model from writing-dna.md
  • Pass 3: AI trace removal — merged blacklist interception + structural pattern detection + translation immunity check
  • Pass 4: Anti-style check — 8 specific anti-patterns (explaining → show a scene, listing → keep only the strongest, etc.)
  • Apply mode-specific overrides (e.g., xiaohongshu allows more emojis; technical-docs adjusts oral test to "clear and concise" standard)
  • Output: Polished draft

Phase 6: Finalize

Run the mode-specific self-check checklist (from the mode file).

  • If all items pass → output final version
  • If any items fail → fix and re-check
  • Output: Final piece ready for publication

Phase Jumping

Users can jump to any phase at any time:

User SaysAction
"重新调研" / "re-research"Go to Phase 1 Step a
"重新找核" / "re-think" / "换个角度"Go to Phase 1 Step c
"改大纲" / "restructure" / "改结构"Go to Phase 2
"重写" / "rewrite"Go to Phase 3
"审校" / "review" / "检查逻辑"Go to Phase 4
"打磨" / "polish" / "去AI味" / "humanize" / "润色"Go to Phase 5
"检查" / "check" / "自检"Go to Phase 6

Each phase is independently re-runnable. Jumping to Phase 4 does not require re-running Phase 1-3.


Tool Integration

This skill can optionally use external tools to enhance quality. Declare these as available capabilities:

CapabilityHow to UseWithout Tools
Web searchUse any available search tool to find current data, competitors, benchmarksGuide user to provide materials
Content extractionUse any available fetch/crawl tool to read URLs and extract contentAsk user to paste content
Academic citationsUse any available academic search tool for papers and citationsMark as "needs verification"
TypesettingInvoke typeset skill if available for rich formattingOutput plain Markdown

No hard dependencies. The skill works fully without any tools — it just works better with them.


Enhanced Capabilities

These modules activate on demand — not loaded unless triggered.

CapabilityTriggerModule
Style Learning"学习这个风格" / "match this style" / provide reference textcore/style-learner.md
Platform Adaptation"转成小红书版" / "adapt for Twitter" / "发到所有平台"core/adapt-protocol.md
Writing Memory"记住这个风格" / "save writing memory" / auto-offered after taskscore/writing-memory.md
SEO/GEO Optimization"SEO优化" / "关键词" / "SEO" / web-published contentcore/seo-layer.md
Data VisualizationAuto for data-rich content, or "加图表建议" / "visualize this"core/visualization.md

These extend the pipeline — they don't replace it. Style learning affects Phase 3+5, SEO adds to Phase 6, visualization enhances Phase 3, etc.


Quick Reference

Writing TypeMode FileBest For
Tech Product Articlemodes/tech-article.mdWeChat, blogs, product launches
Marketing Copymodes/marketing-copy.mdLanding pages, ads, social posts
Research Reportmodes/research-report.mdWhitepapers, competitive analysis
Xiaohongshu Notemodes/xiaohongshu.md种草, tutorials, experience sharing
GitHub READMEmodes/github-readme.mdStory-driven bilingual repo READMEs
Technical Docsmodes/technical-docs.mdAPI docs, changelogs, internal guides
Rewrite / Polishmodes/rewrite.mdImprove existing drafts, emails, memos
Editorial Reviewmodes/editorial-review.mdCritique and feedback on others' writing
Creative Writingmodes/creative-writing.mdFiction, essays, speeches, scripts

评价

还没有评价。欢迎分享你的使用感受。

评论

无需登录即可发表。欢迎分享使用感受、提问,或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。

提交后会立即公开。请勿填写邮箱、电话等个人信息。URL 最多 2 个。

0 / 1000 字

此表单受 reCAPTCHA 保护,适用 Google 的隐私权政策和服务条款。

报告此技能的问题