Two-Layer Prompt Pipeline Architecture

Overview

User-designed architecture for Agentic Media workflow:

WeChat Group Messages
      ↓
News Collector (Agent 1)
      ↓
News Editor (Agent 2) - Kimi 2.6
      ↓
Layout Planner (Agent 3)
      ↓
GPT Image 2 (Agent 4)
      ↓
Daily Newspaper Infographic

Layer 1: Content Editor Agent (Data → Newspaper)

Transforms raw WeChat messages, articles, tweets, GitHub updates into structured news report.

Prompt V1 - AGI Daily Editor: Extracts structured data from raw content:

• 4 most important news items
• 5 most valuable opinions/quotes
• Most important person
• Most important project
• Most important trend

Output: JSON format with headline, hero_story, news[], quotes[], trend{}

Prompt V2 - Deep Editor: Refines content with professional editorial style:

• Wall Street Journal style headlines
• 50-word ledes for each story
• 5 insightful observations
• The Economist / Wired / MIT Technology Review style

Output: Markdown format report

Layer 2: Visual Design Agent (Newspaper → Poster)

Transforms structured content into visual infographic using GPT Image 2.

Layout Planner Output (JSON):

{
  "newspaper_name": "TGO和朋友们的AGI",
  "headline": "AI Agent化浪潮：从工具到同事的进化",
  "hero_image": {
    "concept": "server rack shaking hands with printing press"
  },
  "columns": [
    {"title": "...", "content": "..."}
  ],
  "quotes": [
    {"author": "...", "quote": "..."}
  ]
}

GPT Image 2 Prompt:

• Victorian newspaper style (1880s, The Economist cover)
• Engraving illustration, woodcut etching
• Sepia paper, black ink only
• Main illustration: server rack + printing press (symbolizing AI meets industry)
• CRITICAL: Only include actual content, no fictional sections

Style Composition (User Specified)

The optimal visual style blends:

• 40% New York Times 1900s front page
• 30% The Economist special edition
• 20% US dollar bill engraving style
• 10% Steampunk aesthetic

Key Implementation Notes

1. Content fidelity: Every headline, quote, and summary must come from actual extracted messages
2. No hallucination: GPT Image 2 will invent "International News", "Market Briefing", "1880" dates if not explicitly forbidden
3. Chinese content: All text content is in Chinese; prompt must specify Chinese typography
4. Token efficiency: Keep prompt under 2000 chars for reliable generation