Mediator Skill

Emotional firewall for difficult relationships. Intercepts messages from configured contacts, strips out emotional content, presents just the facts, and helps draft measured responses.

Quick Start

# Initialize config (creates mediator.yaml if missing)
~/clawd/skills/mediator/scripts/mediator.sh init

# Add a contact to mediate
~/clawd/skills/mediator/scripts/mediator.sh add "Ex Partner" \
  --email "ex@email.com" \
  --phone "+15551234567" \
  --channels email,imessage

# Process incoming (usually called by cron/heartbeat)
~/clawd/skills/mediator/scripts/mediator.sh check

# List configured contacts
~/clawd/skills/mediator/scripts/mediator.sh list

# Remove a contact
~/clawd/skills/mediator/scripts/mediator.sh remove "Ex Partner"

Configuration

Config lives at ~/.clawdbot/mediator.yaml:

mediator:
  # Global settings
  archive_originals: true      # Archive raw messages after processing
  notify_channel: telegram     # Where to send summaries (telegram|slack|imessage)
  
  contacts:
    - name: "Ex Partner"
      email: "ex@email.com"
      phone: "+15551234567"
      channels: [email, imessage]
      mode: intercept          # intercept | assist
      summarize: facts-only    # facts-only | neutral | full
      respond: draft           # draft | auto (dangerous)
      
    - name: "Difficult Client"  
      email: "client@company.com"
      channels: [email]
      mode: assist             # Don't hide originals, just help respond
      summarize: neutral
      respond: draft

Modes

• intercept: Archive/hide original, only show summary. User never sees raw emotional content.
• assist: Show original but also provide summary and response suggestions.

Summarize Options

• facts-only: Extract only actionable items, requests, deadlines. No emotion.
• neutral: Rewrite the message in neutral tone, preserving all content.
• full: Show everything but flag emotional/manipulative language.

Respond Options

• draft: Generate suggested response, wait for approval before sending.
• auto: Automatically respond (use with extreme caution).

How It Works

Email Flow

1. Gmail Pub/Sub notification arrives (real-time)
2. Check if sender matches any configured contact
3. If match:
   • Fetch full email content
   • Process through LLM to extract facts/strip emotion
   • Archive original (apply "Mediator/Raw" label, mark read)
   • Send summary to configured notify channel
   • If response needed, draft one

iMessage Flow

1. imsg watch monitors for new messages
2. Check if sender matches configured contact
3. If match:
   • Process message content
   • Send summary to notify channel
   • Draft response if requested

Scripts

• mediator.sh - Main CLI wrapper
• process-email.py - Email processing logic
• process-imessage.py - iMessage processing logic
• summarize.py - LLM-based content analysis and summarization

Integration

Heartbeat Check

Add to HEARTBEAT.md:

## Mediator Check
~/clawd/skills/mediator/scripts/mediator.sh check

Cron (for more frequent checking)

# Check every 5 minutes during business hours
*/5 9-18 * * 1-5 ~/clawd/skills/mediator/scripts/mediator.sh check

Safety Notes

• Never auto-respond to legal, financial, or child-related messages
• Original messages are archived, not deleted (recoverable)
• All actions logged to ~/.clawdbot/logs/mediator.log
• Review and adjust prompts if summaries miss important context

Example Output

Original email:

I can't BELIEVE you would do this to me AGAIN. After everything I've done for you!!! You NEVER think about anyone but yourself. I need you to pick up the kids at 3pm on Saturday and if you can't even do THAT then I don't know what to say anymore.

Mediator summary:

From: Ex Partner Channel: Email
Action Required: Yes

Request: Pick up kids at 3pm Saturday

Suggested response: "Confirmed. I'll pick up the kids at 3pm on Saturday."

See references/prompts.md for the LLM prompts used in processing.