Qiaomu Writer

Qiaomu Writer

Fully automated content creation workflow that produces complete articles in Qiaomu's distinctive style with strategically placed images.

Self-Improving: Writing rules auto-accumulate from your edits. The style gets better every time you revise an AI draft.

Core Philosophy

One-Step Execution: User says "write article with images" → Skill delivers complete article saved to disk.

Failure Resilience: Image generation failures never block article delivery. If MCP fails, save text-only article and report issue.

Zero User Intervention: Complete all steps automatically unless user explicitly interrupts.

Self-Improving Loop: AI draft → You edit → Extract rules → Rules written back to skill → Next draft is more accurate.

Workflow

User Request
    │
    ├─ Research (conditional) → WebSearch
    │
    ├─ Write article → Apply Qiaomu style
    │
    ├─ Plan images → Read image guide, generate prompts
    │
    ├─ Generate images → Call MCP (with error handling)
    │
    ├─ Assemble article → Insert images at strategic positions
    │
    └─ Save to disk → ~/乔木新知识库/10-19 创作/11 公众号-向阳乔木推荐看/YYYY-MM-DD-[slug].md

Step-by-Step Process

Step 1: Research (Conditional)

When to execute:

• User provides topic/keyword only
• User explicitly requests "联网搜索"

When to skip:

• User provides raw content to rewrite
• User says "不用搜索"

Action:

Use WebSearch tool with:
- Query: Expand user's topic into 2-3 search queries
- Focus: Latest information, diverse perspectives
- Limit: Top 10 results sufficient

Step 2: Write Article

Read style guide and anti-tropes first:

Read /Users/joe/.claude/skills/qiaomu-writer/references/writing-style.md
Read /Users/joe/.claude/skills/qiaomu-writer/references/tropes.md

Writing requirements:

1. Qiaomu style characteristics:

   • Conversational tone with "你" (direct address)
   • Short paragraphs with visual breathing room
   • Bold for key points
   • Life analogies for complex concepts
   • Avoid AI-ish phrases like "不是...而是..."
   • Use Chinese commas (，) not dashes

2. Structure:

   • Opening hook (1-2 short paragraphs with question)
   • 3-5 main sections with clear headings
   • Conclusion that echoes opening theme

3. Length:

   • User specifies: Follow user's requirement (e.g., 2000-3000 characters)
   • No specification: Aim for 1500-2500 characters

4. Technical accuracy: Ensure all facts from research are accurate

IMPORTANT: Keep article text in memory, do NOT save yet (images need to be inserted first).

Step 3: Plan Image Strategy

Read image guide first:

Read /Users/joe/.claude/skills/qiaomu-writer/references/image-prompting.md

Selection criteria:

• Opening section: Always include (sets visual tone)
• Core argument sections: 2-3 images (reinforces key points)
• Conclusion: Optional (echoes opening)
• Total: 3-5 images maximum (adjust based on article length)

For each selected section:

1. Extract core concept (one concept, not all details)

2. Choose visual style based on article type:

   • Tech analysis → "赛博朋克" or "商务简约"
   • Product review → "小清新" or "公众号封面"
   • Opinion piece → "水彩插画" or "国潮风格"
   • Tutorial/guide → "商务简约" or "小清新"

3. Construct prompt:

   [seedream模板名称]，主题：[核心概念]，[2-3个视觉元素]，[色调/氛围]
   

4. Record metadata:

   • Section heading (where to insert)
   • Position (after heading)
   • Seedream prompt
   • Image description (5-10 characters for alt text)

Step 4: Generate Images with Error Handling

IMPORTANT: If user said "不需要配图" or "不用配图", skip this step entirely.

For each image prompt:

# Pseudocode for error handling
for each image_prompt in image_prompts:
    try:
        image = mcp__seedream__seedream_text_to_image(
            prompt=image_prompt,
            size="2K",
            watermark=false,
            auto_save=true
        )
        if image.url:
            collected_images.append({
                "section": section_name,
                "url": image.url,
                "alt": alt_text
            })
        else:
            # Image generation returned no URL, skip
            log_warning(f"Image {i} generation returned no URL")
            continue
    except Exception as e:
        # MCP call failed, skip this image
        log_warning(f"Image {i} generation failed: {e}")
        continue

# Continue regardless of success/failure count

Failure handling rules:

• If ALL images fail: Continue with text-only article
• If SOME images fail: Use successful ones
• Never abort article due to image issues
• Report image failures to user at the end

Collect results:

• Seven Niu cloud URLs (preferred for Markdown)
• Or local paths as fallback

Step 5: Assemble Article

Image insertion logic:

For each successfully generated image:

1. Find the corresponding section heading in article text
2. Insert image immediately after the heading
3. Use format: ![alt_text](image_url)
4. Add blank line before and after image for readability

Example:

## 人机协作的新范式

![人机协作](https://newimg.t5t6.com/seedream/...)

你有没有想过，AI工具真正改变的不是你的工作方式...

If no images were generated:

• Use original article text as-is
• Add note in completion report that images were skipped

Step 6: Save Article

File path generation:

~/乔木新知识库/10-19 创作/11 公众号-向阳乔木推荐看/YYYY-MM-DD-[topic-slug].md

Filename rules:

• Use actual current date (YYYY-MM-DD)
• Topic slug: Pinyin or English, simplified (max 30 chars)
• Examples:
  • 2025-11-18-storytelling-skills.md
  • 2025-11-18-ai-programming-tools.md

Create directory if not exists:

mkdir -p ~/乔木新知识库/10-19 创作/11 公众号-向阳乔木推荐看

Save action:

Use Write tool to save the complete article (with or without images)

Step 7: Report Completion and Open Article

Success report format:

✅ 文章已生成并保存

**文件位置**: ~/乔木新知识库/10-19 创作/11 公众号-向阳乔木推荐看/YYYY-MM-DD-[slug].md
**字数**: [count] 字
**配图**: [n] 张图片已插入

[If some images failed:]
⚠️ 图片生成: [success_count]/[total_count] 成功
- 第[x]张图片生成失败（已跳过）

📄 文章预览：
[Show first 200 characters of article]

📖 已在Obsidian中打开文章

Open in Obsidian:

# 自动在Obsidian中打开生成的文章
open "obsidian://open?vault=乔木新知识库&file=10-19 创作/11 公众号-向阳乔木推荐看/YYYY-MM-DD-[slug]"

路径说明：

• 文件保存在知识库的创作区，Obsidian 可直接打开

Usage Examples

Example 1: Full Workflow with Images

User: "联网搜索AI编程工具的发展趋势，然后用乔木风格写一篇文章，并配图"

Execution:

1. WebSearch: "AI编程工具 2025", "GitHub Copilot vs Cursor", "AI代码助手 趋势"
2. Write 1800-character article in Qiaomu style
3. Plan 4 image positions
4. Generate prompts and call MCP for each
5. Successfully generate 3/4 images (one fails, continue)
6. Insert 3 images into article
7. Save to ~/乔木新知识库/10-19 创作/11 公众号-向阳乔木推荐看/2025-11-18-ai-programming-trends.md
8. Report: "✅ 文章已生成，3/4张图片成功"

Example 2: Content Rewrite with Images

User: "把这段内容用乔木风格改写成文章并配图：[text]"

Execution:

1. Skip search (content provided)
2. Rewrite in Qiaomu style
3. Plan 2 image positions
4. Generate and insert images
5. Save article

Example 3: No Images Requested

User: "用乔木风格写一篇关于工作流优化的文章，不需要配图"

Execution:

1. Research topic
2. Write article
3. Skip image generation entirely
4. Save text-only article

Example 4: All Images Fail (Graceful Degradation)

User: "写一篇关于Rust编程的文章并配图"

Execution:

1. Research + Write article
2. Plan 3 images
3. MCP calls all fail (network issue, quota exceeded, etc.)
4. Continue without images
5. Save text-only article
6. Report: "✅ 文章已生成（图片生成失败，已保存纯文本版）"

Error Handling Principles

When MCP Tools Unavailable

If mcp__seedream__* tools not found:
  1. Log warning
  2. Continue with text-only article
  3. Report MCP unavailability in completion message

When Individual Image Generation Fails

For each image failure:
  1. Catch exception
  2. Log which image failed and why
  3. Continue to next image
  4. Never abort article generation

When All Images Fail

If all image generations fail:
  1. Save text-only article
  2. Report image failure count
  3. Suggest checking MCP configuration
  4. Still mark task as "successful" (article delivered)

Quality Standards

Content Quality

• Engaging opening that hooks reader
• Clear logical flow between sections
• Concrete examples over abstract theory
• Natural conversational tone
• No AI clichés or formulaic expressions

Image Quality

• Prompts are specific and actionable
• Visual style matches article tone
• Placement enhances reading flow
• Not too many images (avoid distraction)
• Each image serves a clear purpose

Reliability

• Never abort due to image failures
• Always deliver complete article
• Clear error reporting
• Predictable file paths

Important Notes

• Full automation: Execute all steps without user confirmation
• Error resilience: Image failures never block article delivery
• User experience first: Completed article is the priority, images are enhancement
• Clear reporting: Always tell user what succeeded and what failed

Self-Improving Mode

Trigger phrases: "提取写作规则", "我改了这篇文章", "更新写作风格", "帮我从这次修改中提取规则"

When to use: After you edit an AI-generated article to your satisfaction, feed both versions back to extract style rules. Rules accumulate automatically in writing-style.md — zero manual writing.

Workflow

Step 1: Get both files
  - AI draft: the file Claude saved originally
  - Revised: your edited version (or paste the content)

Step 2: Run diff + extract rules
  bash /Users/joe/.claude/skills/qiaomu-writer/scripts/extract-style-rules.sh <draft.md> <revised.md>
  → Generates a Claude prompt with the diff embedded

Step 3: Paste that prompt to Claude
  → Claude analyzes what you changed and extracts reusable rules

Step 4: Review extracted rules
  → You decide which rules are worth keeping (takes 30 seconds)

Step 5: Append approved rules to writing-style.md
  → Claude appends to: references/writing-style.md

Rule format (auto-generated)

## 自动积累的规则（YYYY-MM-DD）

### 规则 N：[规则标题]
- **现象**：AI 写了什么（坏的）
- **修改**：乔帮主改成了什么（好的）
- **原则**：从这次修改提炼的通用规则

Manual shortcut (in-conversation)

If you paste an AI draft + your revision directly in chat, Claude can extract rules inline without the script:

"这是AI初稿和我的修改版，帮我提取写作规则追加到 writing-style.md"

Stats tracking

After appending rules, report:

• Total rules in writing-style.md (count ### 规则 headers)
• Rules added this session
• Date of last update

References

• references/writing-style.md - Qiaomu style guide (auto-accumulates rules over time)
• references/tropes.md - AI writing anti-tropes checklist (30+ patterns to avoid)
• references/image-prompting.md - Image strategy and prompt engineering
• scripts/extract-style-rules.sh - Self-improving: diff two article versions → extract rules