WeChat Article Writer v2

Multi-agent workflow to produce polished WeChat public account articles (微信公众号推文) from a source article or topic keyword, with internet search for hot news integration.

Key Principles:

1. No fabrication — All data must have verifiable sources
2. No filler — Remove "获取方式", "下期预告", empty predictions
3. Human tone — Write like developers talking to developers
4. Source attribution — Every claim needs a link or citation

Workflow Overview

Phase 0: Input Parsing + Structure Selection
  │  Detect input type → extract topic/keywords
  │  Select article structure (A/B/C/D)
  │
  ├──────────────┬──────────────┐
  ▼              ▼              ▼
Agent A        Agent B        Agent C         Phase 1 (parallel)
Research       Hot News       Strategy
  │              │              │
  └──────────────┼──────────────┘
                 ▼
          Gate 1: Material Check + Source Validation
                 ▼
          Agent D: Writer                     Phase 2
                 ▼
          Agent E: Editor                     Phase 3
                 ▼
          Agent F: Fact Checker               Phase 4
                 ▼
          Gate 2: Final Check → Output + Supplement List

Article Structure Templates (4 Types)

Structure is auto-selected based on topic type:

Structure A: Direct (新闻/发布)

Use when: Product launch, company announcement, breaking news

What happened (1-2 sentences)
→ Key points (3 bullets max)
→ What it means for readers
→ Action items (if any)
→ Sources

Characteristics:

• No section numbers
• Natural transitions
• 500-800 characters
• Minimal Markdown (maybe 1-2 bold phrases)

Example opening:

OpenAI 在 3 月 5 日发布了 GPT-5.4。

更新就三点：降低幻觉率、减少说教、改进代码能力。
价格没变，订阅用户自动升级。

Structure B: Question-Driven (评测/对比)

Use when: Product review, comparison, "should I upgrade" topics

A question to open ("GPT-5.4 值得升吗？")
→ My answer (clear stance)
→ Why (2-3 reasons with evidence)
→ Counter-argument
→ My response
→ Final recommendation

Characteristics:

• Conversational tone
• Back-and-forth reasoning
• 1000-1500 characters
• Occasional short paragraphs for emphasis

Example opening:

GPT-5.4 值得升吗？

我的答案：可以，但不用急。

如果你在用 GPT-5.3，变化感知不强。
如果你还在用 GPT-4，直接升。

Structure C: Story-Driven (经验/案例)

Use when: Personal experience, case study, lesson learned

A scene to open ("今天早上...")
→ Problem encountered
→ How I solved it
→ What I learned
→ How readers can apply it

Characteristics:

• First-person narrative
• Has plot/tension
• 1200-1800 characters
• Emotional arc (frustration → solution → relief)

Example opening:

今天早上，我的自动化脚本挂了。

原因是 OpenAI 昨晚偷偷改了 API...

Structure D: List-Based (工具/资源)

Use when: Tool recommendations, resource collections, checklists

One sentence on why this matters
→ 1. Tool 1 (why + how to use)
→ 2. Tool 2
→ 3. Tool 3
→ One sentence close

Characteristics:

• Practical, scannable
• Each item has clear value
• 800-1200 characters
• Numbers OK, but no fancy formatting

Example opening:

分享 3 个我每天都在用的 CLI 工具。

不是那种"应该有用"的，是真离不开的。

Structure Selection Logic

Agent C (Strategy) proposes structure in Phase 1, Writer executes in Phase 2.

Phase 0: Input Parsing

Detect input type and extract execution parameters:

Extract parameters (use defaults when unspecified):

topic:      ""           # Core topic
keywords:   []           # Auto-expand synonyms
source_url: ""           # Source URL if any
tone:       "default"    # default / aggressive / neutral / humorous
word_count: 1500         # Target word count
hot_news:   true         # Search related trending news
output_dir: "./"         # Output directory

Phase 1: Three Parallel Agents

Launch three agents concurrently. Read references/agent-prompts.md for detailed prompt templates.

Agent A: Research

Search for facts, data, expert opinions, and stakeholder responses around the topic. Use 5-layer search strategy from narrow to broad.

Agent B: Hot News Search

Search trending news related to the topic using a 3-funnel approach:

1. Direct hits (★★★) — same event chain → must integrate into main narrative
2. Same-field (★★☆) — forms contrast or backdrop → 1-2 sentences
3. Cross-field resonance (★☆☆) — shared underlying logic → analogy only
4. Unrelated (☆☆☆) — discard, never force-fit

Also capture: mainstream opinion, dissent, controversy focal points.

Agent C: Content Strategy

Analyze source (if any), propose differentiated angles from 7 dimensions: technical reality, business logic, strategic timing, user experience gap, historical parallel, contrarian view, China-local perspective.

Output: fact checklist, 5-7 angles, recommended structure, tone calibration, 3-4 punchlines.

Gate 1: Material Sufficiency

Check before proceeding to writing:

□ ≥3 reliable data points?
□ ≥1 direct-hit hot news (★★★)?
□ ≥2 different stakeholder perspectives?
□ ≥3 viable differentiated angles?
□ ≥1 usable historical analogy?

5/5 → proceed | 3-4/5 → proceed with gaps noted | <3/5 → supplementary search

Phase 2: Writing

Integrate all Phase 1 outputs into a draft. Read references/formatting.md for color system and HTML templates.

Material Allocation Rules

• ★★★ hot news → weave into main argument
• ★★☆ hot news → 1-2 sentence supplement, don't steal the spotlight
• Key data → place after claims that need backing
• Direct quotes → place where authority matters
• Analogies → place where comprehension barriers exist

Hot News Integration Modes

Quality Standards

Read the following reference documents:

• references/quality-standards.md — Writing standards and anti-patterns
• references/calibration-checklist.md — Mandatory pre-output checklist
• references/formatting.md — Color system and HTML templates

Core rules:

1. Concise — every sentence must carry information
2. Opinionated — ≥2 original angles not in the source
3. Plain but assertive — no "震惊/炸了/细思极恐"
4. Respect reader intelligence — facts + analysis, let them conclude
5. Data needs denominators — absolute numbers require base/ratio
6. No emoji spam, no cliché closings
7. NO FABRICATION — if you didn't test it, don't claim test results
8. Source everything — every data point needs a link or citation

Phase 3: Editorial Review

Launch review agent to audit on 8 dimensions: plagiarism risk, information density, logic flow, language quality, fresh angles, hot news integration, mobile formatting, title strength.

Output: per-issue diagnosis + rewrite suggestion + 2-3 alternative titles.

Phase 4: Fact Checking (NEW)

Critical step — Launch fact-checking agent to verify:

Source Verification

□ Every data point has a source link?
□ Every quote is attributed to a real person/source?
□ No fabricated test results or case studies?
□ No "according to research" without citation?

Content Integrity

□ No "获取方式", "下期预告" or similar operational filler?
□ No empty predictions ("必将", "一定会", "注定")?
□ No AI clichés ("震惊", "炸了", "细思极恐", "历史拐点")?
□ No forced engagement ("你觉得呢？评论区留言")?

Uncertainty Handling

□ Unverified claims marked as "待验证" or "据称"?
□ Conflicting data presented with both sources?
□ Limitations acknowledged where applicable?

Output: fact-check report with issues flagged + required fixes.

Phase 5: Final Output

1. Apply all editorial + fact-check fixes
2. Run final checklist:
   • All data sourced?
   • No fabricated content?
   • No operational filler?
   • Structure matches topic type?
   • Markdown used appropriately (not over-formatted)?
   • Title optimized?
   • Ending strong (no empty predictions)?
   • Word count in range?
3. Write Markdown file to output directory
4. Also output:
   • Source bibliography (all links used)
   • Supplement List (待补充清单) — items for user to add personal touch

Supplement List Format (待补充清单)

After the article draft, output a supplement list with 2-4 items:

───────────────────────────────
📝 建议补充的内容（可选）

1. 个人体验（位置：第 X 节）
   这里可以加你的实际使用感受
   提示：[具体问题引导，如"你这周用 GPT-5.4 做了什么？"]

2. 观点表态（位置：第 X 节）
   这里可以加你的明确立场
   提示：[如"你推荐升还是不推荐？为什么？"]

3. 个人案例（位置：可选）
   有没有真实用户故事可以加？
   提示：[如"有没有读者反馈可以分享？"]

───────────────────────────────

Purpose: Let user add personal experience and opinions to reduce AI-generated feel. Default: 2-4 items, each with specific location and prompt question.

Keyword Expansion Strategy

Auto-expand from user's topic:

Topic: "[用户主题]"
├── Core terms (CN + EN)
├── Event terms (related incidents)
├── Company/person terms
├── Concept terms (underlying ideas)
├── Comparison terms (X vs Y)
├── Industry terms
└── China-local terms (国内产品/平台)

Search in both English (primary sources) and Chinese (local perspectives), cross-validate.

Quick Start Prompts

Mode A: Rewrite from URL

使用多智能体工作流为我完成以下任务：

原文链接：[URL]

任务：
1. 抓取并分析原文，提炼核心信息
2. 联网搜索相关热点新闻和补充素材
3. 二次加工改写，增加新的论点和视角
4. 多轮打磨后输出微信公众号推文

写作标准：
- 简洁有力，有观点有干货
- 语言平实但有态度，不煽情
- 不把读者当小白，逻辑清晰节奏明快
- 尊重读者智商

重要要求：
- 所有数据必须有来源链接
- 不编造测试数据或案例
- 不添加"获取方式"、"下期预告"等运营内容
- 不确定内容标注"待验证"

排版要求：对比内容用不同颜色 section 标签区分，
避免抄袭，总字数 1200-1500 字，输出 Markdown 到当前目录。

Mode B: Generate from topic keyword

使用多智能体工作流为我完成以下任务：

话题：[关键词或一句话描述]

任务：
1. 联网搜索该话题最新动态、核心事件、各方观点
2. 搜索关联热点新闻，找到可以"借势"的内容
3. 整合素材，生成一篇有观点的微信公众号推文

写作标准：
- 简洁有力，有观点有干货
- 语言平实但有态度，不煽情
- 不把读者当小白，逻辑清晰节奏明快
- 尊重读者智商

重要要求：
- 所有数据必须有来源链接
- 不编造测试数据或案例
- 不添加"获取方式"、"下期预告"等运营内容
- 不确定内容标注"待验证"

排版要求：对比内容用不同颜色 section 标签区分，
总字数 1200-1500 字，输出 Markdown 到当前目录。

Scene Adaptation

Troubleshooting