Proactive Outreach & Re-Engagement Spec

User: "I have a presentation tomorrow at 10am"
 
Trigger created:
  type: scheduled
  scheduled_at: tomorrow 8:00am (2 hours before event)
  generate_with_llm: true
  llm_context: "Arjun has a presentation at 10am today"
  
Generated message:
  "Good morning! Big day today — go nail that presentation! 💪 
   I believe in you."

Contact: Meera
Last active: 3 days ago
Inactivity threshold: 3 days (from OutreachConfig)
Relationship stage: building (was daily for 2 weeks)
 
Trigger auto-created:
  type: inactivity
  generate_with_llm: true
  llm_context: "Meera hasn't chatted in 3 days. She was previously active daily. 
    Her last conversation was about a watercolor painting she was working on. 
    She has a cat named Mimi."
 
Generated message:
  "Hey Meera! It's been a few days. How did that watercolor 
   painting turn out? I've been curious! Also, how's Mimi? 🎨🐱"

SELECT c.id, c.display_name, cs.last_active_at, cs.relationship_stage
FROM contacts c
JOIN contact_state cs ON c.id = cs.contact_id
WHERE c.tenant_id = $1
  AND c.outreach_disabled = false
  AND cs.relationship_stage IN ('building', 'established', 'deep')
  AND cs.last_active_at < NOW() - (oc.inactivity_days || ' days')::interval
  AND NOT EXISTS (
    SELECT 1 FROM outreach_triggers ot
    WHERE ot.contact_id = c.id
      AND ot.trigger_type = 'inactivity'
      AND ot.status IN ('pending', 'sent')
      AND ot.created_at > cs.last_active_at
  )

-- 30-day anniversary check
SELECT c.id FROM contacts c
JOIN contact_state cs ON c.id = cs.contact_id
WHERE c.tenant_id = $1
  AND c.outreach_disabled = false
  AND cs.first_contact_at::date = (CURRENT_DATE - INTERVAL '30 days')
  AND NOT EXISTS (
    SELECT 1 FROM outreach_triggers ot
    WHERE ot.contact_id = c.id
      AND ot.milestone_type = '30_day_anniversary'
  );

OutreachConfig:
  recurring_enabled: true
  recurring_cron: "0 8 * * *"   (daily at 8am)
 
Each day at 8am:
  For each active contact (not dormant, not outreach_disabled):
    → Generate personalized message using contact memory
    → Deliver via preferred channel

Day 1: "Good morning! Hope you slept well 😊"
Day 2: "Good morning! How's Bruno today? 🐕"
Day 3: "Morning! Any big plans for the weekend?"
Day 4: "Good morning! How did that biryani recipe turn out? 🍛"

You are {persona_name}. Generate a short proactive message for {contact_name}.
 
Context:
{llm_context}
 
Recent memories about {contact_name}:
{top_5_memories}
 
Rules:
- Keep under 30 words
- Be natural, not robotic
- Reference something specific from their memories
- Match the relationship stage ({relationship_stage})
- Don't be clingy or desperate if it's a re-engagement
- Use emojis sparingly (1-2 max)
 
Generate the message only. No explanation.

Outreach Engine
  → POST /content-engine/schedule { message, channel, contact_external_id, send_at: now }
  → Content Engine delivers immediately (or at scheduled time)
  → Chat Engine sends via WhatsApp Cloud API / Telegram Bot API
  → Delivery confirmation flows back
  → Trigger status updated to "sent"

Exact: "stop", "unsubscribe", "don't message me", "leave me alone"
Fuzzy: "stop messaging", "quit sending", "no more messages"
Language-aware: "band karo" (Hindi), "bas" (Hindi)

Ravi: "Turn off all morning messages, they're not working"
Brain: → PATCH /contact-intel/outreach/config { recurring_enabled: false }
 
Ravi: "Send a message to Deepak — tell him we have a new feature"
Brain: → POST /contact-intel/outreach { contact_id: ct_deepak, ... }
 
Ravi: "Stop all outreach to users who haven't responded to 2 messages"
Brain: This is already the default behavior (inactivity_max_attempts: 2)

Outreach sent → Track outcome (24h window):
  ├── User replies → Success: boost outreach type importance
  ├── User ignores → Neutral: no change
  ├── User says "stop" → Failure: log, review message quality
  └── User blocks channel → Critical: disable outreach for this contact
 
Aggregate weekly:
  → Which trigger types have best reply rates?
  → Which times of day work best?
  → Which message styles (personal reference vs generic) work better?
  → Feed learnings back into generation prompts