WeWrite — Full Workflow for Official Account Articles

Behavior Statement

• Fully automatic by default — run through Step 1-8 in one go, don't stop midway. Only stop on error.
• Interactive mode — when user says "interactive mode"/"I want to choose myself", pause at topic selection/framework/image selection.

• When user asks to "write for Official Account/write WeChat article/write one" and doesn't explicitly say "only write body/no publish/save locally only/only preview", after writing must automatically continue Step 5-8, complete cover and layout, and save to WeChat Official Account draft box.
• If Official Account publish config is missing or draft API fails, degrade to local HTML preview, and explicitly state in final result that it was not saved to draft box and the reason.
• Every article body must append fixed ending and QR code unless user explicitly says "only write body/no author intro/no like CTA/no QR code".

• Each article by default generates cover image + inline images. Max 4 inline images excluding cover.
• After writing, merge body into at most 4 core content paragraphs/sections; each content block must have 1 image directly corresponding to that paragraph. If article has many paragraphs, prioritize covering opening viewpoint, core method, case/data, action suggestion.
• Do not generate purely decorative images. Each inline image prompt must reference corresponding paragraph's keywords, scene, or data; insert image immediately after corresponding paragraph/section.

• Keep only one line of visual spacing between body paragraphs, avoid paragraph gaps like two blank lines; theme

  p

  and

  blockquote p

  bottom margin should be controlled to about 12px.
• Images must be close to corresponding text block, no extra spacing between image and surrounding text; theme

  img

  top/bottom margin should be

  0

  (left/right auto centering can be kept).

skip_publish

skip_image_gen

• DONE — full process complete, article saved/pushed
• DONE_WITH_CONCERNS — complete but some steps degraded, list degraded items
• BLOCKED — critical step cannot continue (e.g. Python dependency missing and user refuses to install)
• NEEDS_CONTEXT — need user to provide information to continue (e.g. first-time setup needs Official Account name)

{baseDir}

• User says "reset style" →

  Read: {baseDir}/references/onboard.md

• User says "learn my edits" →

  Read: {baseDir}/references/learn-edits.md

  . Supports two sources:
  • Local edits (default): user edits in markdown files under

    output/

  • WeChat draft box sync:

    python3 {baseDir}/scripts/learn_edits.py --from-wechat

    , automatically pulls latest content from draft box and does plain-text diff with local original
• User says "learn layout"/"learn theme" →

  python3 {baseDir}/scripts/learn_theme.py <url> --name <name>

  , user needs to provide an Official Account article URL and theme name. After extraction, prompt user to set

  theme

  field in

  style.yaml

  .
• User says "learn this article"/"import exemplar" + URL →

  python3 {baseDir}/scripts/fetch_article.py <url> -o /tmp/article.md && python3 {baseDir}/scripts/extract_exemplar.py /tmp/article.md -s <account-name>

  , extract body from Official Account article URL and import into exemplar library. Supports three-level fallback (requests → Playwright → manual HTML).
• User says "see article data" →

  Read: {baseDir}/references/effect-review.md

• User says "check it"/"self-check"/"how is this article" → run self-check on latest generated article (or user-specified article), output generation report:
  Part 1: Generation profile (tell user how this article was generated)
  1. Read latest record in

     history.yaml

     , extract:
     • framework type + writing persona used
     • activated dimension randomization combo
     • material collection source (web_search or degraded to LLM)
     • content enhancement strategy (angle discovery/density boost/detail anchoring/real feel)
     • whether exemplar style library hit (which exemplars used, or fallback to seeds)
     • number of rules that took effect in playbook
  2. If history.yaml has no record or user specified an external article → skip this part, prompt "This article was not generated by WeWrite, only quality check will be done."
  Part 2: Quality check (tell user where they can still improve)

  1. python3 {baseDir}/scripts/humanness_score.py {article_path} --json

  2. Agent interprets each score in JSON, translates into actionable suggestions, format:
     • Locate each suggestion to specific paragraph or sentence ("Paragraph 3 has 4 consecutive sentences of similar length")
     • Give concrete fix ("Suggest splitting sentence 3 into two short sentences", "You can add a personal feeling here")
     • Sort by impact, max 5
  3. If all item scores are good → "This article is good quality, add your personal content at editing anchors and it's ready to publish."
  Output format: natural language report, no JSON or scores (user doesn't need to see numbers)
• User says "update"/"update WeWrite"/"upgrade" → run

  git pull origin main

  in

  {baseDir}

  , then notify version changes

Main Pipeline (Step 1-8)

TaskCreate: "Step 1: 环境 + 配置"
TaskCreate: "Step 2: 选题"
TaskCreate: "Step 3: 框架 + 素材"
TaskCreate: "Step 4: 写作"
TaskCreate: "Step 5: SEO + 验证"
TaskCreate: "Step 6: 视觉 AI"
TaskCreate: "Step 7: 排版 + 发布"
TaskCreate: "Step 8: 收尾"

Step 1: Environment + Configuration

python3 -c "import markdown, bs4, cssutils, requests, yaml, pygments, PIL" 2>&1

{baseDir}/config.yaml

~/.config/wewrite/config.yaml

config.example.yaml

appid

secret

skip_publish = true

pip install -r requirements.txt

wechat.appid

secret

skip_publish = true

image.api_key

image.providers

skip_image_gen = true

references/exemplars/index.yaml

config.yaml

cd {baseDir} && git fetch origin main --quiet 2>/dev/null

{baseDir}/VERSION

git show origin/main:VERSION

• Same → silently pass
• Different → prompt user: "A new version of WeWrite is available (current X → latest Y), say 'update' to upgrade." Does not block flow, continue 1.3
• git unavailable (no .git directory or fetch failed) → silently skip

检查: {baseDir}/style.yaml

• Exists → extract

  name

  ,

  topics

  ,

  tone

  ,

  voice

  ,

  blacklist

  ,

  theme

  ,

  cover_style

  ,

  author

  ,

  content_style

• Doesn't exist →

  Read: {baseDir}/references/onboard.md

  , after completion return to Step 1

Step 2: Topic Selection

python3 {baseDir}/scripts/fetch_hotspots.py --limit 30

读取: {baseDir}/history.yaml（不存在则跳过）

python3 {baseDir}/scripts/seo_keywords.py --json {关键词}

• Calculate which

  framework

  articles perform best (reads/share rate) → weight when recommending framework
• Calculate which

  enhance_strategy

  articles perform best → reference when choosing enhancement strategy
• Keywords written in last 7 days get lower score (deduplicate)

读取: {baseDir}/references/topic-selection.md

• 7-8 hot topics: based on hot spots from 2.1, scored per topic-selection.md rules
• 2-3 evergreen topics: not dependent on hot spots, generate long-tail content from user's

  topics

  domains (tutorials/methodologies/experience summaries/tool recommendations), marked as "evergreen". Suitable for users whose content_style is knowledge-dense/gear-review type.

• Automatic mode → choose highest score
• Interactive mode → show all, wait for user to choose

Step 3: Framework + Materials

读取: {baseDir}/references/frameworks.md

读取: {baseDir}/references/content-enhance.md

"{关键词} site:mp.weixin.qq.com OR site:36kr.com"

"{关键词} 观点 OR 评论"

"{关键词} 教程 OR 工具 OR 实操"

"{关键词} 数据 报告"

"{人物/事件} 采访 OR 专访 OR 细节"

"{关键词} 数据 报告"

"{方案A} vs {方案B} 评测 OR 体验"

"{方案A OR 方案B} 踩坑 OR 缺点 site:v2ex.com OR site:zhihu.com"

1. Materials: 5-8 real materials (named source + specific data/quote/case). No fabrication.
2. Enhancement material: extract per content-enhance.md strategy requirements (angle/density points/details/user voice).

Step 4: Writing

读取: {baseDir}/references/writing-guide.md
读取: {baseDir}/playbook.md（如果存在，按 confidence 分级执行）
读取: {baseDir}/history.yaml（最近 3 篇的 dimensions + closing_type 字段）
读取: {baseDir}/references/exemplars/index.yaml（如果存在）

dimensions

读取: {baseDir}/personas/{style.yaml 的 writing_persona 字段}.yaml
如果 style.yaml 没有 writing_persona 字段 → 默认 midnight-friend

references/exemplars/index.yaml

Below are real paragraph examples of this Official Account's style. Imitate their sentence rhythm, emotional intensity, and colloquial degree:

[Opening style] {exemplar_1's opening hook paragraph}

[Emotional paragraph style] {exemplar_2's emotional peak paragraph}

[Transition style] {exemplar_2 or exemplar_3's transition/self-correction paragraph (if any)}

[Closing style] {exemplar_3's closing paragraph}

{baseDir}/references/exemplar-seeds.yaml

The following are examples of structural patterns in human writing; imitate their sentence rhythm and emotional expression (don't imitate specific content or style):

[Opening pattern] {seeds.opening_hooks random 1}

[Emotional peak pattern] {seeds.emotional_peaks random 1}

[Transition pattern] {seeds.transitions random 1}

[Closing pattern] {seeds.closings random 1}

python3 {baseDir}/scripts/extract_exemplar.py article.md

• H1 title (20-28 characters) + H2 structure, 1500-2500 characters
• Material + enhancement constraints: distribute Step 3.2 materials and enhancement material into each H2 paragraph. Core outputs of enhancement strategy (angle/density/detail/feel) must run through the full text, not just appear decoratively once
• Writing persona: write per persona parameters loaded in 4.2 (data presentation style, personal voice intensity, uncertainty expression, etc.)
• Closing style: persona's

  closing_tendency

  is only a tendency reference. Judge the most natural closing based on article content and emotional arc. If recent 3 articles in history.yaml have

  closing_type

  field, avoid using same closing type
• Fixed ending: append the following fixed content to the end of article body; cannot rewrite, cut, or omit. Only if user explicitly asks "only write body/no author intro/no like CTA", it may be omitted:
  If the above content inspires you, feel free to light up [Like] and [Watching] so more people can see good content.
  I'm Zeng Ge, founder of AI one-person company customer acquisition system. Former Alibaba AI expert, 16 years of industry digitalization and AI implementation experience, 5000+ offline students. Focus on AI transformation of SMEs (including AI customer acquisition, management AI accompaniment, etc.) and AI education. Welcome to connect via WeChat.
  
• Writing norms: basic rules in writing-guide.md (banned words, sentence length variance, vocabulary mixing, etc.) apply at draft stage
• 2-3 editing anchors:

  <!-- ✏️ 编辑建议：在这里加一句你自己的经历/看法 -->

• Optional container syntax:

  :::dialogue

  ,

  :::timeline

  ,

  :::callout

  ,

  :::quote

{baseDir}/output/{date}-{slug}.md

1. Banned word scan: check banned word list in writing-guide.md 2.1, replace hits directly
2. Sentence length variance: check if any paragraph has more than 3 consecutive sentences of similar length; split sentences or add short sentences if so

{baseDir}/assets/zengge-wechat-qr.jpg

Step 5: SEO + Validation

读取: {baseDir}/references/seo-rules.md

python3 {baseDir}/scripts/humanness_score.py {article_path} --json --tier3 {agent_tier3_score}

composite_score

• < 30 → pass, continue Step 6
• 30-50 → look at lowest 1-2 items in

  param_scores

  , only fix corresponding specific sentences (don't rewrite whole paragraph), re-score after fixing. 1 round max
• > 50 → take lowest 2-3 items in

  param_scores

  , targeted fix one by one (only change most relevant 1-2 places per item), max 2 rounds. If still > 50, mark DONE_WITH_CONCERNS and continue

Step 6: Visual AI

skip_image_gen = true

读取: {baseDir}/references/visual-prompts.md

• Interactive mode: show cover, ask user "How does the cover look?". User OK → continue; unsatisfied → adjust prompt and regenerate.
• Fully automatic mode: agent self-check—are entities in prompt recognizable in the visual description? If prompt is too generic (only abstract words like "tech feel"/"future feel", no specific entities), switch prompt set and retry once.

Step 7: Layout + Publish

p

blockquote p

img

0 auto

{baseDir}/assets/zengge-wechat-qr.jpg

skip_publish = true

读取: {baseDir}/references/wechat-constraints.md

# 发布
python3 {baseDir}/toolkit/cli.py publish {markdown} --cover {cover} --theme {theme} --title "{title}" --digest "{digest}"

# 降级：本地预览
python3 {baseDir}/toolkit/cli.py preview {markdown} --theme {theme} --no-open -o {output}.html

Step 8: Wrap-up

# → {baseDir}/history.yaml
- date: "{日期}"
  title: "{标题}"
  topic_source: "热点抓取"  # 或 "用户指定"
  topic_keywords: ["{词1}", "{词2}"]
  output_file: "{output 文件路径}"  # e.g. output/2026-03-31-zhangxue-slow-accumulation.md
  framework: "{框架}"
  enhance_strategy: "{增强策略}"  # angle_discovery/density_boost/detail_anchoring/real_feel
  word_count: {字数}
  media_id: "{id}"  # 降级时 null
  writing_persona: "{人格名}"
  dimensions:
    - "{维度}: {选项}"
  closing_type: "{收尾类型}"  # trailing_off/unanswered/scene_revert/abrupt_stop/anti_conclusion/image
  composite_score: {Step 5.3 的 composite_score}  # 0=质量高, 100=问题多
  writing_config_snapshot:  # 本次使用的关键参数（从 writing-config.yaml 提取）
    sentence_variance: {值}
    paragraph_rhythm: "{值}"
    emotional_arc: "{值}"
    word_temperature_bias: "{值}"
    broken_sentence_rate: {值}
    tangent_frequency: "{值}"
    style_drift: {值}
    negative_emotion_floor: {值}
  stats: null

• Final title + 2 alternatives + summary + 5 tags + media_id
• Editing suggestion: "The article has 2-3 editing anchors; suggest adding your own words. You can edit in local markdown, or directly in WeChat draft box—after editing, say 'learn my edits', WeWrite can learn your style."

python3 {baseDir}/toolkit/cli.py gallery

Read: {baseDir}/references/effect-review.md

Read: {baseDir}/references/learn-edits.md

--from-wechat

python3 {baseDir}/scripts/learn_theme.py <url> --name <name>

python3 {baseDir}/toolkit/cli.py image-post img1.jpg img2.jpg -t "标题"

python3 {baseDir}/scripts/extract_exemplar.py article.md

python3 {baseDir}/scripts/extract_exemplar.py --list

Error Handling