Multi-Agent Content Pipeline with Claude Code

.claude/
└── skills/
    └── discover-pioneers/
        ├── index.md          # Main orchestrator
        └── prompts/
            ├── research.md   # Research agent instructions
            └── write.md      # Write agent instructions

# index.md frontmatter
---
name: discover-pioneers
description: Find and profile people who build personal AI operating systems
---

Research → Photo → Write → Humanizer → SEO
   ↓         ↓        ↓         ↓        ↓
 JSON     Image   Markdown   Clean    Optimized
 data     file     draft     prose    article

# Research Agent Prompt

Find a person who:
1. Documents how they personally use AI (not just builds tools)
2. Has a systematic approach to AI-augmented work
3. Shares their methodology publicly

## Candidate Scoring

| Criteria | Points |
|----------|--------|
| Documents personal daily AI workflow | +6 |
| Has systematic methodology (named system) | +5 |
| Builds tools for personal use | +4 |
| Writes about AI philosophy | +3 |
| Active content about personal AI | +3 |
| Open-sources configs (CLAUDE.md, prompts) | +2 |

1. Check GitHub avatar, Twitter, personal site
2. Download to static/images/people/{slug}.jpg
3. Verify it shows a real human face (use image analysis)
4. If cartoon/logo, delete and try next source
5. If no photo found, continue without it

## Article Structure

{Opening: Who is this person and what's unique? No header.}

## Background
- Previous work
- Current role
- What led them to build their system

## The System
**Daily workflow:** How they use AI day-to-day
**Core tools:** Table of what they use and how
**Key automations:** Specific things they've built

## Why This Matters
2-3 paragraphs of analysis. NOT bullet points.

## What You Can Steal
| Technique | How to Apply |
Actionable takeaways readers can copy.

## Already Written (EXCLUDE these)

{{existing_people}}

# Pseudocode for what the orchestrator does
existing = []
for f in Path('content/people/').glob('*.md'):
    frontmatter = parse_frontmatter(f)
    existing.append(frontmatter['title'])

research_prompt = research_template.replace(
    '{{existing_people}}', 
    '\n'.join(existing)
)

from sentence_transformers import SentenceTransformer

model = SentenceTransformer('all-MiniLM-L6-v2')

def index_article(path: str):
    content = read_markdown(path)
    embedding = model.encode(content)
    db.execute(
        "INSERT INTO documents (path, content, embedding) VALUES (?, ?, ?)",
        [path, content, embedding.tobytes()]
    )

def search(query: str, limit: int = 5):
    query_vec = model.encode(query)
    # SQLite doesn't do vector similarity natively
    # Use numpy or a vector extension
    results = []
    for row in db.execute("SELECT path, content, embedding FROM documents"):
        doc_vec = np.frombuffer(row['embedding'])
        similarity = cosine_similarity(query_vec, doc_vec)
        results.append((similarity, row))
    return sorted(results, reverse=True)[:limit]

# In Clawdbot config
cron:
  - schedule: "0 9 * * *"  # 9am daily
    task: "/skill discover-pioneers"

# Generate 5 articles
for i in {1..5}; do
  claude "/skill discover-pioneers"
  sleep 60  # Rate limiting
done

.claude/skills/
└── my-content-pipeline/
    ├── index.md
    └── prompts/
        └── research.md

---
name: my-content-pipeline
description: Generate content about [your topic]
---

## Process

1. Research: Launch agent with prompts/research.md
2. Write: Use research output to draft article
3. Review: Check for patterns to remove
4. Save: Write to content/[section]/

## Research prompt injection

{{existing_articles}}