human-tone-polish

name

human-tone-polish

description

Polish Chinese technical, educational, and public-facing drafts to reduce mechanical or template-like AI tone while preserving the author's voice, technical accuracy, inline citations, figures, captions, and article structure. Use when the user asks to 去 AI 味、人味儿润色、打磨文章、精简啰嗦段落、改掉模板话术、优化技术文章语气、保留作者风格、润色正稿或润色稿，including but not limited to WeChat public-account, blog, course note, research explainer, and MagicAI/魔方AI空间 article workflows.

Human Tone Polish

Purpose

Polish Chinese technical and explanatory drafts so they read like the author's own article, not a generic AI rewrite.

Prioritize, in order:

1. Preserve technical correctness.
2. Preserve the author's position and phrasing habits.
3. Remove mechanical rhythm, template transitions, and over-explained prose.
4. Improve public-reading rhythm.
5. Keep citations, image placeholders, figure numbers, and tables coherent.

Workflow

1. Read the draft and nearby context.

   • Inspect the target article first.
   • If available, inspect the outline, collaboration draft, final draft, user-edited version, image prompt file, or reference article in the same topic folder.
   • Treat the latest user-edited draft as the strongest style signal.

2. Decide how much to touch.

   • For a requested section polish, edit only that section unless the surrounding transition clearly breaks.
   • For a whole-article polish, keep the structure unless the draft has duplicated or obviously slow sections.
   • Prefer small, visible improvements over rewriting every paragraph.

3. Apply the human-polish pass.

   • Use references/rewrite-principles.md for editing decisions.
   • Keep direct technical sentences when they already work.
   • Remove writing-process phrases and empty scaffolding.
   • Shorten repeated examples, caveats, and conclusion padding.

4. Apply the final check.

   • Use references/ai-tone-checklist.md before delivery.
   • Check headings, image numbering, captions, links, and Markdown rendering.
   • Run a lightweight diff or Markdown check when files are edited.

5. Learn from useful feedback.

   • When the user flags a phrase, rhythm, heading, or rewrite as AI-like, decide whether the feedback is reusable beyond the current article.
   • If it is reusable, update the relevant reference file in this skill during the same turn.
   • Store style rules, not one-off facts. Do not add article-specific names, claims, links, or temporary preferences as permanent rules.
   • Keep each new rule concrete enough to guide a future edit.

Editing Rules

• Do not make the prose uniformly polished. Some asymmetry is author voice.
• Do not turn every section into the same pattern.
• Do not add invented scenes, emotions, or personal anecdotes.
• Do not remove a caveat if the remaining technical claim becomes misleading.
• Do not rewrite a clear definition just to make it sound more stylish.
• Do not add slogans such as 未来可期, 全面赋能, 革命性突破, or 生态闭环 unless the author explicitly wants that tone.

Output

When editing files:

• Create -润色稿.md when polishing an existing -正稿.md, unless the user asks to overwrite.
• Preserve existing image references and placeholders.
• Briefly summarize the main changes and any claims that still need source verification.

When giving inline suggestions:

• Provide the improved sentence or paragraph directly.
• Explain only the key reason, such as 更短, 少一点模板感, 技术边界更稳.