feat: add humanization layer to remove AI writing patterns

Integrate de-AI writing system based on blader/humanizer (29 AI patterns)
and Chinese-specific rules (去AI味). Voice calibration via writing samples
lets the system match the user's personal style.

- lib/humanize.md: shared de-AI rule library (29 patterns + 中文专项)
- content/config/voice.yaml: voice calibration template
- skills/media-write: add humanization audit pass (Step 5) before variants
- skills/media-setup: add voice configuration wizard (Step 5)
- CLAUDE.md: document humanization feature

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: pianweifen

CLAUDE.md

@@ -42,6 +42,14 @@ A set of Claude Code skills that let developers write content once and publish t
 
 **v2 (planned):** 小红书, 知乎, Medium
 
+## Humanization (de-AI writing)
+
+All content passes through a humanization layer defined in `lib/humanize.md`:
+- **29 AI writing patterns** detected and rewritten (based on Wikipedia's WikiProject AI Cleanup)
+- **Chinese-specific de-AI rules** (去AI味): anti-成语堆砌, 口语化, structural de-rigidification
+- **Voice calibration** via `content/config/voice.yaml` — provide writing samples to match personal style
+- Applied during drafting (rules active while writing) and as a dedicated audit pass before variant generation
+
 ## Image generation
 
 Configurable backend via `IMAGE_PROVIDER` in `.env`:

content/config/voice.yaml

@@ -0,0 +1,29 @@
+# Voice Calibration Config
+# Provide writing samples so the system matches your personal style.
+# This takes priority over generic de-AI rules in lib/humanize.md.
+
+# Paste 2-3 paragraphs of your own writing here.
+# The system will extract voice traits automatically.
+samples:
+  - |
+    # Paste your first writing sample here.
+    # Blog post, README, tweet thread — anything in your natural voice.
+  - |
+    # (Optional) A second sample for better calibration.
+
+# Or manually specify traits (auto-extracted if samples are provided):
+traits:
+  sentence_length: medium      # short | medium | long
+  vocabulary: conversational   # academic | technical | conversational | casual
+  humor: occasional            # none | dry | occasional | frequent
+  perspective: first_person    # first_person | second_person | third_person
+  formality: informal          # formal | neutral | informal
+
+# Per-platform tone overrides (optional):
+# platform_overrides:
+#   devto:
+#     formality: informal
+#     humor: frequent
+#   wechat:
+#     vocabulary: conversational
+#     perspective: first_person

lib/humanize.md

@@ -0,0 +1,124 @@
+# Humanize — De-AI Writing Guide
+
+Shared library for removing AI-generated writing patterns. Read this file
+before writing or revising any content.
+
+Based on Wikipedia's WikiProject AI Cleanup patterns and Chinese content
+de-AI techniques.
+
+## Core Principle
+
+LLMs produce text that trends toward the most statistically likely phrasing.
+Human writing is specific, uneven, opinionated, and occasionally messy.
+The goal is not to trick detectors — it is to write things a real person
+would actually say.
+
+## The 29 AI Writing Patterns to Avoid
+
+### Content Patterns
+
+| # | Pattern | Example | Fix |
+|---|---------|---------|-----|
+| 1 | Significance inflation | "marking a pivotal moment" | State the fact plainly |
+| 2 | Notability name-dropping | "featured in Forbes, TechCrunch..." | Only cite if directly relevant |
+| 3 | Superficial -ing analyses | "symbolizing, reflecting, showcasing" | Say what it actually does |
+| 4 | Promotional language | "breathtaking", "game-changing" | Remove or replace with specifics |
+| 5 | Vague attributions | "Experts believe", "Many developers say" | Name the source or drop it |
+| 6 | Formulaic challenges | "Despite challenges... continues to" | Be specific about what went wrong |
+
+### Language Patterns
+
+| # | Pattern | Example | Fix |
+|---|---------|---------|-----|
+| 7 | AI vocabulary | "delve, leverage, landscape, utilize, foster, tapestry, multifaceted, nuanced, comprehensive, robust" | Use plain words: dig into, use, field, area |
+| 8 | Copula avoidance | "serves as", "stands as", "features", "boasts" | Just say "is" or "has" |
+| 9 | Negative parallelisms | "It's not just X, it's Y" | Pick one point and make it |
+| 10 | Rule of three | Always listing exactly 3 items | Vary list lengths: 2, 4, 5 |
+| 11 | Synonym cycling | Repeating ideas with different words | Say it once, move on |
+| 12 | False ranges | "from architecture to deployment" | Be specific about what you cover |
+| 13 | Passive voice / subjectless fragments | "No configuration needed" | Name the actor: "You don't need to configure anything" |
+
+### Style Patterns
+
+| # | Pattern | Example | Fix |
+|---|---------|---------|-----|
+| 14 | Em dash overuse | "the tool — which is free — works" | Use commas or split into two sentences |
+| 15 | Boldface overuse | Random **emphasis** everywhere | Bold only for genuinely key terms |
+| 16 | Inline-header lists | "**Label:** content" format | Use normal prose or plain lists |
+| 17 | Title Case Headings | "How To Build A Great API" | Sentence case: "How to build a great API" |
+| 18 | Decorative emojis | "Let's get started! :rocket:" | Remove unless platform convention |
+| 19 | Curly quotes | Typographic quotes in code contexts | Use straight quotes |
+| 20 | Hyphenated buzzwords | "cross-functional, data-driven, future-proof" | Unpack: say what you mean |
+| 21 | Persuasive authority tropes | "At its core", "What truly matters" | Drop it; start with the point |
+| 22 | Signposting announcements | "Let's dive in", "Here's what you need to know" | Just start |
+| 23 | Fragmented headers | Headers that need body text to make sense | Headers should stand alone |
+
+### Communication Patterns
+
+| # | Pattern | Example | Fix |
+|---|---------|---------|-----|
+| 24 | Chatbot artifacts | "I hope this helps! Let me know if..." | Remove entirely |
+| 25 | Cutoff disclaimers | "While details are limited..." | Remove or be specific about what's unknown |
+| 26 | Sycophantic tone | "Great question!", "Absolutely!" | Don't grade the reader |
+
+### Filler & Hedging
+
+| # | Pattern | Example | Fix |
+|---|---------|---------|-----|
+| 27 | Filler phrases | "In order to", "It's worth noting that" | "To", or just state it |
+| 28 | Excessive hedging | "could potentially possibly help" | One qualifier max |
+| 29 | Generic conclusions | "The future looks bright", "Only time will tell" | End with a concrete point or don't conclude |
+
+## Chinese Content (中文去AI味)
+
+When writing in Chinese, also watch for these patterns:
+
+### 结构问题
+- **过度结构化**: 不要用"首先、其次、最后"、"第一、第二、第三"等机械排列。真人写作结构更松散，偶尔跳跃。
+- **小标题 + emoji 排列**: 每段都带 emoji 小标题是典型 AI 格式，减少使用。
+- **万能总结句**: "综上所述"、"总而言之"、"总的来说" — 删掉或换成具体观点。
+
+### 词汇问题
+- **成语堆砌**: AI 爱用四字成语显得"有文化"，真人日常写作成语密度低得多。
+- **高频 AI 词**: "赋能、驱动、生态、闭环、底层逻辑、颗粒度、抓手" — 用大白话替代。
+- **过度修饰**: "深入浅出地探讨"、"全方位多角度" — 删掉修饰，直说。
+- **重复连接词**: "不仅……而且"、"无论……都" — 减少使用，换成口语转折。
+
+### 语气问题
+- **过于客观**: AI 写作缺少主观判断。加入"我觉得"、"说实话"、"有点意外的是"等个人视角。
+- **过于流畅**: 真人写作有停顿、口语化表达、偶尔的不完整句。适当加入语气词。
+- **缺少具体例子**: AI 喜欢泛泛而谈。用具体的数字、日期、项目名、个人经历替代概括性描述。
+
+### 修复策略
+1. 用第一人称，加入主观感受
+2. 打破工整结构，允许段落长短不一
+3. 用口语替代书面语（"搞定"代替"完成"，"踩坑"代替"遇到问题"）
+4. 举具体例子，带时间和细节
+5. 偶尔用反问、自嘲、吐槽增加真实感
+
+## Voice Calibration
+
+If `content/config/voice.yaml` exists, read it before writing. It contains:
+- Writing samples from the user
+- Extracted voice traits (sentence length, vocabulary level, humor style, etc.)
+- Per-platform tone adjustments
+
+When voice.yaml is present, match the user's natural style rather than
+applying generic humanization. The user's own patterns take priority
+over the rules above.
+
+## How to Apply
+
+This guide is used in two places:
+
+### During drafting (per-section)
+Apply these rules as you write each section. Don't draft in AI-speak
+and fix later — write naturally from the start.
+
+### Audit pass (after full draft)
+After the complete source.md is written, do a dedicated scan:
+1. Read the full text looking for each of the 29 patterns
+2. Flag any instances found
+3. Rewrite flagged passages
+4. If the content is in Chinese, also apply the Chinese-specific checks
+5. Show the user what changed and why

skills/media-setup/SKILL.md

@@ -124,7 +124,77 @@ RESPONSE=$(curl -s -o /dev/null -w "%{http_code}" \
 [ "$RESPONSE" = "200" ] && echo "OK: OpenAI connected" || echo "ERROR: Auth failed"
 ```
 
-## Step 5: Write Config Files
+## Step 5: Configure Writing Voice (Optional)
+
+Ask the user if they want to set up their personal writing style.
+
+Use AskUserQuestion:
+"Want to set up your writing voice? This helps the system match your personal style instead of sounding like AI."
+- A) Yes, paste writing samples
+- B) Yes, describe my style manually
+- C) Skip for now
+
+### Option A: Writing samples
+
+Ask the user to paste 2-3 paragraphs of their own writing (blog posts, READMEs, tweets, etc.).
+
+Use AskUserQuestion:
+"Paste a writing sample — anything you've written that sounds like *you*. Blog post, README, email, tweet thread, 公众号文章 — all work. (Paste one now, I'll ask for more after.)"
+
+Collect 1-3 samples. For each, extract voice traits:
+- Average sentence length (short/medium/long)
+- Vocabulary level (academic/technical/conversational/casual)
+- Humor presence (none/dry/occasional/frequent)
+- Default perspective (first_person/second_person/third_person)
+- Formality (formal/neutral/informal)
+- Language patterns (e.g., uses rhetorical questions, starts with anecdotes, uses 口语化 Chinese)
+
+Show the extracted traits to the user for confirmation:
+```
+Based on your samples, here's your voice profile:
+  sentence_length: short
+  vocabulary: conversational
+  humor: dry
+  perspective: first_person
+  formality: informal
+
+Does this look right? I can adjust any of these.
+```
+
+### Option B: Manual style description
+
+Use AskUserQuestion to walk through each trait:
+
+1. "How long are your typical sentences?" (short and punchy / medium / long and flowing)
+2. "What's your vocabulary like?" (academic / technical jargon / conversational / casual slang)
+3. "Do you use humor in your writing?" (never / dry/subtle / sometimes / a lot)
+4. "What perspective do you write from?" (I/我 first person / you/你 second person / third person)
+5. "How formal is your tone?" (formal / neutral / informal/casual)
+
+### Write voice.yaml
+
+Write the voice config to `content/config/voice.yaml`:
+
+```yaml
+# Generated by /media-setup
+samples:
+  - |
+    <pasted sample 1>
+  - |
+    <pasted sample 2>
+
+traits:
+  sentence_length: <extracted or chosen>
+  vocabulary: <extracted or chosen>
+  humor: <extracted or chosen>
+  perspective: <extracted or chosen>
+  formality: <extracted or chosen>
+```
+
+Tell the user:
+"Voice profile saved. The writing skills will match this style automatically. You can update it anytime by editing `content/config/voice.yaml` or running `/media-setup` again."
+
+## Step 6: Write Config Files
 
 Write credentials to `.env` (gitignored):
 ```bash
@@ -153,7 +223,7 @@ platforms:
   - github-pages
 ```
 
-## Step 6: Configure git-lfs (Optional)
+## Step 7: Configure git-lfs (Optional)
 
 If the user expects to publish many posts with images:
 
@@ -169,7 +239,7 @@ git lfs track "content/**/assets/*.jpg"
 git add .gitattributes
 ```
 
-## Step 7: Summary
+## Step 8: Summary
 
 Print a summary of what was configured:
 
@@ -184,6 +254,9 @@ Platforms:
 Image generation:
   [OK] OpenAI DALL-E 3
 
+Writing voice:
+  [OK] Voice profile configured (informal, conversational, first person)
+
 Next steps:
   - Run /media-idea to brainstorm your first post
   - Run /media to start the full guided workflow

skills/media-write/SKILL.md

@@ -24,6 +24,7 @@ Write content collaboratively, then generate platform-specific variants.
 Before starting, read these shared library files:
 - Read `lib/adapter-discovery.md` for adapter format.md location
 - Read `lib/manifest-ops.md` for manifest creation
+- Read `lib/humanize.md` for de-AI writing rules (MUST read before any writing)
 
 ## Step 1: Get the Brief
 
@@ -74,11 +75,14 @@ Work through the article section by section:
 
 ### Writing guidelines
 
+- **Apply `lib/humanize.md` rules from the start.** Do not draft in AI-speak and fix later.
 - Write in the user's voice (if brand voice is configured in `content/config/voice.yaml`, read it first)
 - Include code examples where relevant for a developer audience
 - Use clear headers (## for sections, ### for subsections)
 - Keep paragraphs concise (3-5 sentences)
 - Add image placeholders where visuals would help: `![description](assets/placeholder.png)`
+- Avoid the 29 AI patterns listed in `lib/humanize.md` — no significance inflation, no AI vocabulary, no filler phrases
+- For Chinese content, follow the 中文去AI味 section: use 口语化 tone, avoid 成语堆砌, add personal perspective
 
 ## Step 4: Finalize Source
 
@@ -89,7 +93,38 @@ Use AskUserQuestion:
 - A) Yes, generate variants
 - B) I want to edit more (specify which section)
 
-## Step 5: Generate Platform Variants
+## Step 5: Humanization Audit
+
+After the user approves the full draft, run a dedicated de-AI pass before generating variants.
+
+### Process
+
+1. Re-read `lib/humanize.md` to refresh the pattern list.
+2. Read the complete `source.md`.
+3. Scan for each of the 29 AI writing patterns. For Chinese content, also apply the 中文去AI味 checks.
+4. For each flagged passage, rewrite it following the "Fix" column in the pattern table.
+5. Show the user a summary of changes:
+   ```
+   Humanization audit — <N> patterns found and fixed:
+     - Section 2: removed significance inflation ("pivotal moment" → "useful")
+     - Section 4: replaced AI vocabulary ("leverage" → "use")
+     - Section 5: removed filler phrase ("It's worth noting that")
+   ```
+6. Use AskUserQuestion:
+   "Here are the humanization edits. Accept all, or review individually?"
+   - A) Accept all
+   - B) Review each change
+   - C) Skip humanization (keep original)
+7. Apply accepted changes to `source.md`.
+
+### Voice calibration (optional)
+
+If `content/config/voice.yaml` exists:
+- Compare the draft against the user's writing samples
+- Adjust sentence length, vocabulary level, and tone to match
+- Voice.yaml traits take priority over generic humanization rules
+
+## Step 6: Generate Platform Variants
 
 For each configured platform adapter:
 
@@ -106,7 +141,7 @@ For each configured platform adapter:
 
 3. Write the variant to `<post_dir>/variants/<name>.md`
 
-## Step 6: Create Manifest
+## Step 7: Create Manifest
 
 Create `manifest.yaml` following the format in `lib/manifest-ops.md`:
 
@@ -131,7 +166,7 @@ variants:
 
 Use the atomic write pattern from `lib/manifest-ops.md`.
 
-## Step 7: Summary
+## Step 8: Summary
 
 ```
 Article written and variants generated: