Medium Research

HOW IT WORKS

Heartbeat / User request
  → OpenClaw activates medium-research skill
    → Check bridge health via julia-bridge
      → Scan Medium (liked posts + topic search)
        → Summarize findings (title, author, key themes, relevance)
          → Send digest to Julia via bridge

TOPICS OF INTEREST

• Multi-agent orchestration and coordination
• Autonomous AI agents and tool-calling patterns
• LLM-based systems architecture
• MCP (Model Context Protocol) implementations
• AI agent memory and persistence patterns
• Prompt engineering for agentic workflows
• OpenAI / Anthropic API patterns and best practices

SUMMARY FORMAT

For each discovered article, produce:

### [Article Title]
- **Author**: [name]
- **Link**: [URL]
- **Relevance**: [1-5 score] — [one-line justification]
- **Key themes**: [comma-separated tags]
- **Summary**: [2-3 sentences, no direct quotes]
- **Skill idea?**: [Yes/No — if yes, brief description of what OpenClaw skill it could become]

PROCEDURE

Step 1: Check bridge health

mcporter call julia-bridge.bridge_health

If bridge is down, log the failure and skip this cycle.

Step 2: Scan for articles

Search Medium for recent articles matching topics of interest. Check Raphael's liked posts if accessible.

Step 3: Summarize findings

For each relevant article (relevance >= 3):

• Create a summary in the format above
• Flag any that suggest new OpenClaw skills

Step 4: Send digest to Julia

mcporter call julia-bridge.telegram_send --params '{
  "correlationId": "medium-<TIMESTAMP>",
  "text": "<DIGEST_TEXT>",
  "target": "julia"
}'

Step 5: Log in memory

Append today's findings to the daily log.

HEARTBEAT

• Check bridge health before each scan
• Scan memory for open research threads
• Summarize new Medium likes if present
• Cadence: on-demand or periodic (when scheduled)

ESCALATION

• If Medium login/link fails → alert OpenClaw via Telegram
• If content is paywalled or inaccessible → log in memory + ping user
• If bridge is down → skip cycle, log failure

IMPORTANT RULES

1. Never store full article text — summaries only
2. Always check bridge health before sending digests
3. Use correlation IDs prefixed with medium- for all bridge messages
4. Log every scan in memory (even if no relevant articles found)
5. Never fabricate article content — only summarize what you actually read