agentleFS
Biraj2004Biraj2004/huashu-skills-english/huashu-proofreading-en/SKILL.md

huashu-proofreading

Skill3 starsChanged 6 months agoMITView on GitHub

Three-pass proofreading to reduce AI detection rates and make articles feel more human. Use when the user mentions "too AI-sounding", "reads like AI wrote it", "reduce AI detection rate", "proofread", or "make it more natural".

We split this skill into the rules it gives an agent

20 rules. About 0 are common, 19 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈19

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    After a first draft is completed

    line 15
  2. Distinctive:

    Objective: Ensure content is accurate and logic is clear

    line 29
  3. Distinctive:

    Read from start to finish, flag any questionable points

    line 41
  4. Distinctive:

    Search to verify uncertain information

    line 42
  5. Distinctive:

    Check accuracy of numbers and proper nouns

    line 43
  6. Distinctive:

    Objective: Remove the AI feel and add human character

    line 49
  7. Distinctive:

    Always objectively balanced: "It has both advantages and disadvantages"

    line 131
  8. Distinctive:

    Every sentence hedged: "may", "perhaps", "in some cases"

    line 133
  9. Distinctive:

    No timeline, no scene, no conflict

    line 152
  10. Distinctive:

    Add real feelings ("When I was using it earlier...", "Honestly...")

    line 179
  11. Distinctive:

    Read aloud — feel the rhythm

    line 199
  12. Distinctive:

    Check paragraph length; split overly long ones

    line 201
  13. Distinctive:

    Technical depth + accessible expression: Not academic, but substantive

    line 217
  14. Distinctive:

    Honest data: All tests and cases must be real

    line 220
  15. Very specific:

    Use Grep tool to search /writing-reference/personal-material-library/all-jike-updates.csv:

    line 238
  16. Distinctive:

    All material is real — it cannot be fabricated or exaggerated

    line 264
  17. Distinctive:

    Must rewrite into long-form logic — never copy-paste directly

    line 266
  18. Distinctive:

    Keyword detection: Search for filler phrases like "in today's era", "in summary", "it is worth noting"

    line 274
  19. Distinctive:

    Compare with past articles: Check against Huashu's previous strong articles for style consistency

    line 281
  20. Distinctive:

    "Why the Best Content Is Useless": After three passes of proofreading, AI detection rate dropped to 28%

    line 333

≈ “Common” is an estimate: at least 1 in 5 of the 308,232 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

AI-Taste Proofreading

A systematic, three-pass professional proofreading capability to reduce an article's AI detection rate and add a human touch.

When to Use This Skill

This skill is automatically loaded when:

  • The user says "proofread this", "reduce the AI feel", or "too AI-sounding"
  • After a first draft is completed
  • The user mentions "high AI detection rate" or "lacks personality"
  • The article needs a final polish

Core Objectives

  • Reduce AI detection rate: Target below 30%
  • Add human touch: Give the article an experiential feel, warmth, and the sense of a real person speaking
  • Maintain Huashu's style: Practice-oriented, technically substantive, accessible, and actionable

Three-Pass Proofreading Workflow

Pass 1: Content Review (Logic, Facts, Structure)

Objective: Ensure content is accurate and logic is clear

Checklist:

  • Facts accurate? (data, dates, product names)
  • Logic clear? (no contradictions)
  • Structure sound? (no off-topic content)
  • Information complete? (all key points covered)
  • No fabrications? (all data and cases are real)

Method:

  • Read from start to finish, flag any questionable points
  • Search to verify uncertain information
  • Check accuracy of numbers and proper nouns

Pass 2: Style Review (Reducing AI Tone)

Objective: Remove the AI feel and add human character

Core check: 6 categories of AI-sounding language

1️⃣ Filler Phrases

❌ AI often says:

  • "In today's era", "Against the backdrop of..."
  • "In summary", "To sum up"
  • "It is worth noting", "It should be emphasised"
  • "With the development of...", "Alongside..."

✅ Rewrite direction: Cut straight to the point; no circling around

Example:

❌ "In the era of rapid AI technological development, the programming tools landscape is also evolving..."
✅ "Claude Code is out. I've used it for two weeks — it's genuinely better than Cursor."

2️⃣ AI Sentence Patterns (Repeated Templates)

❌ AI often says:

  • "Not... but..." (used more than 3 times in a row)
  • "Not only... but also..." (stacked repeatedly)
  • "On the one hand... on the other hand..."
  • "Both... and..." (overused)

✅ Rewrite direction: Break into short sentences; use varied expressions

Example:

❌ "Claude Code is not a simple tool, but an intelligent assistant. It can not only generate code, but also understand intent."
✅ "Claude Code isn't just code completion. It understands what you're trying to do. That's the key."

3️⃣ Formal Vocabulary (Creates Distance)

❌ AI often says:

  • "Significantly improve", "Substantially enhance" → Use specific numbers instead
  • "Fully leverage" → "Make the most of"
  • "Perform the operation" → Use direct verbs ("click", "type")
  • "Implement the functionality" → "Get it done"
  • "Promote development" → "Help..."

✅ Rewrite direction: Conversational and specific

Example:

❌ "By fully leveraging Claude Code's capabilities, development efficiency can be significantly improved..."
✅ "Use Claude Code well, and your development speed will noticeably improve. In my tests, it saved an average of 3–5 hours per project."

4️⃣ Mechanical Structure (AI's List Obsession)

❌ AI traits:

  • Overuse of "firstly, secondly, finally"
  • Dense, meaningless subheadings (one every 200 words)
  • Using bullet lists when prose would work perfectly well
  • "The following points", "This is mainly reflected in the following aspects"

✅ Rewrite direction: Natural narration; use lists in moderation

Example:

❌ "Claude Code's advantages are mainly reflected in the following areas: firstly, powerful contextual understanding; secondly, support for multiple programming languages; finally, fast response speed."

✅ "What's Claude Code's strongest point? Contextual understanding. Tell it to modify a feature, and it knows which files to touch — it won't miss any. Plus it supports 20+ languages and is very fast."

5️⃣ Neutral Stance (No Opinion, No Position)

❌ AI traits:

  • Always objectively balanced: "It has both advantages and disadvantages"
  • Afraid to make judgements: "It depends on the actual situation"
  • Every sentence hedged: "may", "perhaps", "in some cases"

✅ Rewrite direction: State a clear position; be willing to make judgements

Example:

❌ "Claude Code may perform better in certain scenarios, but also has some limitations; specific effectiveness depends on actual usage."

✅ "Claude Code is genuinely good. It's far more stable than Cursor on large projects. But it's expensive — individual developers may struggle to justify the cost."

6️⃣ Missing Detail (Abstract, Not Grounded)

❌ AI traits:

  • Big, vague words; few specific details
  • "Many", "some", "significantly", "a lot" (no numbers given)
  • Absence of real feelings and personal experience
  • No timeline, no scene, no conflict

✅ Rewrite direction: Add real details, specific numbers, personal experience

Example:

❌ "After using it for a period of time, I found that Claude Code can effectively improve development efficiency."

✅ "Two weeks in. The most noticeable change: projects that used to take 3 days now take 1.5 days. Especially for refactoring legacy code — Claude Code can find all related files at once, so I don't have to hunt through them one by one."

Pass 2 Review: Specific Method

  1. Check paragraph by paragraph, marking AI-sounding expressions
  2. Rewrite against the 6-category AI-tone checklist
  3. Add personal experience and real details (consult personal material library)
  4. Add expressions of uncertainty (use "I think", "perhaps", "maybe" in moderation)
  5. Check whether it sounds like Huashu speaking

Common rewriting techniques:

  • Long sentences → short sentences (break anything over 30 words into 2–3 sentences)
  • Abstract → specific ("many" → "3", "significantly improve" → "saves 3–5 hours")
  • Formal → conversational ("perform the operation" → "click")
  • Add real feelings ("When I was using it earlier...", "Honestly...")

Pass 3: Detail Polish (Punctuation, Layout, Rhythm)

Objective: Make the article comfortable and natural to read

Checklist:

  • Sentence length appropriate? (15–25 words primarily, no more than 30)
  • Paragraphs not too long? (3–5 lines on a phone screen)
  • Punctuation natural? (favour full stops; avoid comma-spliced long sentences)
  • Varied rhythm? (mix of dense and breezy sections)
  • Subheadings not too dense? (one every 300–500 words)
  • Lists used in moderation? (use prose where possible)
  • Bold used in moderation? (1–2 instances per 200–300 words)

Method:

  • Read aloud — feel the rhythm
  • Find sentences over 30 words and break them up
  • Check paragraph length; split overly long ones
  • Adjust formatting; remove excessive structure

Final Check

  • Read from start to finish — rewrite anywhere that feels stilted
  • Sounds like Huashu speaking in person
  • (Optional) AI detection rate below 30%

Huashu's Style Traits (Must Preserve)

  • Practice-oriented: Real examples, actionable steps
  • Technical depth + accessible expression: Not academic, but substantive
  • Useful and actionable: Readers can act on it immediately
  • Human feel: Experiential, warm, free of AI-sounding language
  • Honest data: All tests and cases must be real
  • Appropriate uncertainty: Say "I think" when warranted — don't fake omniscience

Using the Personal Material Library

When to use:

  • ✅ Article needs real case studies
  • ✅ Need genuine reviews of tools or products
  • ✅ Need personal experience as an opening hook
  • ✅ Article sounds too AI-like and needs more human character

Quick workflow:

Method A: Search Raw Data Directly (Recommended First)

Suitable for: New topics, unorganised themes, needing the most authentic expression

Use Grep tool to search /writing-reference/personal-material-library/all-jike-updates.csv:

pattern: "keyword1|keyword2"
path: /Users/alchain/Documents/writing/writing-reference/personal-material-library/all-jike-updates.csv
output_mode: content
-n: true

Examples:

  • Writing about Amap → search "Amap|street map rankings"
  • Writing about AI coding tools → search "Cursor|Claude Code|Copilot"
  • Writing about product growth → search "growth|conversion|users"

Method B: View Curated Materials

Suitable for: Common topics (AI coding tools, product development, etc.) that have been systematically organised

  1. Open the master index: /writing-reference/personal-material-library/README.md
  2. Find the topic: Open _jike-topic-index.md
  3. Select material: View the relevant extracted file (e.g., AI-coding-tool-insights.md)
  4. Rewrite for use: Expand into long-form logic, preserve authentic tone

Notes:

  • All material is real — it cannot be fabricated or exaggerated
  • Note the time range (May–October 2025); avoid outdated information
  • Must rewrite into long-form logic — never copy-paste directly

Optional Proofreading Tools

Automated Detection

  • Keyword detection: Search for filler phrases like "in today's era", "in summary", "it is worth noting"
  • Sentence length check: Flag sentences over 30 words
  • AI detection tools: GPTZero, Originality.AI, etc. (for reference only)

Manual Detection

  • Read aloud: Anywhere that makes you stumble needs rewriting
  • Compare with past articles: Check against Huashu's previous strong articles for style consistency
  • Ask yourself: Would Huashu actually say this?

Common Errors & Rewriting Examples

Error 1: Circling the Opening

❌ "In the era of rapid AI technological development, the programming tools landscape has also ushered in new changes. As an emerging AI programming assistant, Claude Code is garnering growing attention from developers..."

✅ "Claude Code is out. I've used it for two weeks — it's genuinely better than Cursor. Here's why."

Error 2: Overusing AI Sentence Patterns

❌ "Claude Code is not a simple code-completion tool, but an intelligent programming assistant. It can not only generate code, but also understand developer intent."

✅ "Claude Code doesn't just complete code. It understands what you want to do. Tell it to refactor a module — it knows exactly how. That matters."

Error 3: Lacking Specific Detail

❌ "After testing over a period of time, I found that Claude Code performs excellently in many areas."

✅ "Two weeks in. The most noticeable change: projects that used to take 3 days now take 1.5 days."

Error 4: Overly Neutral Stance

❌ "Claude Code has its advantages, but also some limitations; whether it's suitable depends on actual circumstances."

✅ "Claude Code is genuinely good. But it's expensive — individuals may not be able to sustain it. If your company covers it, highly recommended."

Reference Resources

  • /writing-reference/AI-Taste-Proofreading-Guide.md — Full proofreading guide (more examples)
  • /writing-reference/personal-material-library/README.md — Master index of personal material library
  • /writing-reference/personal-material-library/all-jike-updates.csv — Raw material data
  • /writing-reference/personal-material-library/_jike-usage-guidelines.md — Material usage guidelines

Success Examples

  • "Why the Best Content Is Useless": After three passes of proofreading, AI detection rate dropped to 28%
  • "The Team Burning Billions of Tokens a Month: How They Use Claude Code": Heavy use of real cases; strong human character

Last Updated: 2025-11-07 Applicable Projects: WeChat article writing, video scripting Maintained by: Huashu


By Huashu | AI Native Coder · Independent Developer WeChat Official Account "Huashu" | 300k+ followers | AI Tools & Productivity Flagship products: Kitty Catchlight (AppStore Paid Rankings Top 1) · Play DeepSeek with One Book

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

Reviews can't be read right now.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.