Build Your Own 24/7 AI Assistant: ClawdBot Replica with Claude Code

0

Mindset Shift: Build vs. Wait

FOUNDATION

ClawdBot triggered everyone's imagination about what's possible — proactive AI, 24/7 availability, voice calls. But 42,665 instances were exposed with critical vulnerabilities (CVSS 9.6 severity, 93.4% auth bypass rate), three CVEs were filed, and API costs hit $500-5,000/month. Google's VP of Security Engineering called it "infostealer malware disguised as an AI personal assistant." Instead of waiting for someone else to fix it, you take what you like and bring it to your own system. This is the builder's mindset.

❌ Before (Waiting)

• Hoping ClawdBot security gets patched
• Non-technical users vibe-coding security (false safety)
• $500-5,000/month variable API bills
• 42,665 exposed instances — CVSS 9.6, 93.4% auth bypass
• 3 CVEs filed (command injection, unauth access, arbitrary file access)
• Someone else controls your AI infrastructure

✅ After (Building)

• You own the entire stack
• $200/month fixed on Claude Max plan
• Runs on your machine — no exposed instances
• 2-hour rebuild (if you already use Claude Code)
• Sky's the limit on customization

📋 DECISION CHECKLIST

Watch GodaGoo's video to understand what's being built (link)

Decide: Are you willing to keep a computer running 24/7? (Laptop or VPS)

Subscribe to Claude Max plan ($200/month) if not already

Install Claude Code CLI on your dedicated machine

Verify Node.js 18+ is installed (node --version)

⚠️ The Experts Were Unequivocal:
"Infostealer malware disguised as an AI personal assistant." — Heather Adkins, VP Security Engineering, Google Cloud
"My current favorite for the most likely Challenger disaster in coding agent security." — Simon Willison, AI Researcher
"22% of enterprise customers have employees actively using Clawdbot. Shadow AI risk is real." — Token Security Research

🔒 Full Security Comparison: ClawdBot vs. Claude Code

Vulnerability	ClawdBot	Claude Code
Network Exposure	Public WebSocket :18789	Local only, no ports
Credential Storage	Plaintext JSON files	MCP OAuth, env vars
Authentication	None by default	User ID restriction
Execution Model	Auto-execute all	Permission-gated
Sandboxing	None	Claude Code sandbox

CVEs: CVE-2025-6514 (command injection), CVE-2025-49596 (unauthenticated access), CVE-2025-52882 (arbitrary file access). 26% of 31,000 ClawdBot skills had vulnerabilities.

📊 Feature Comparison: ClawdBot vs. Claude Code Always-On

Feature	ClawdBot	Claude Code
24/7 Operation	✓	✓
Voice Messages	✓	✓
Phone Calls	✗	✓
Semantic Memory	✗	✓
Security	42K exposed	Private
Your Infrastructure	✗	✓

🔮 Builder's Insight: GodaGoo had everything else already built in Claude Code. The only NEW features he needed were 24/7 availability and Telegram connection. If you already use Claude Code with skills and MCP servers, you're 80% there. The playbook just bridges the last 20%.

🤖 AI Prompt: Assess Your Starting Point

Use this prompt to have Claude Code audit your current setup and identify what you already have vs. what you need to build.

I want to build a 24/7 AI assistant accessible via Telegram with:
1. Always-on Claude Code (headless)
2. Telegram bot relay for messaging
3. Persistent memory (Supabase or local)
4. Bidirectional voice calling (11 Labs + Twilio)
5. Proactive check-ins (calendar, email, tasks)
6. Observability dashboard

Audit my current Claude Code setup:
- What skills/MCP servers do I already have?
- What integrations exist (Google, Notion, etc.)?
- What's my OS and environment?

Then create a GAP ANALYSIS:
- What I already have (list with ✅)
- What I need to build (list with ❌)
- Estimated time per gap
- Recommended build order

Output as a structured checklist I can follow.

📊 Cost Comparison: ClawdBot vs. Build Your Own

ClawdBot (Hosted)

• Heartbeat only: ~$150/month
• Active use: $500-5,000/month
• Opus 4.5 API burns fast
• Variable, unpredictable
• Security: 42K exposed instances

Your Own (Claude Code)

• Claude Max: $200/month (fixed)
• Supabase: Free tier
• 11 Labs + Twilio: $11-20/month
• Total: ~$250/month fixed
• Security: Your machine only

Mark Step 0 Complete

↑ Return to Top

1

Set Up Claude Code Always-On (Headless)

CORE

Claude Code needs to run 24/7 in the background without a visible terminal window. This "headless" mode keeps the AI engine always available to receive and process requests from Telegram, voice calls, and proactive check-in schedules.

❌ Before

• Claude Code only runs when terminal is open
• Close laptop lid = assistant goes dark
• No way for external services to trigger it
• Manual start every session

✅ After

• Claude Code runs as background service
• Survives terminal close and sleep settings
• Listens for incoming requests continuously
• Auto-restarts if process crashes

📋 RECIPE

Disable sleep/hibernate on your dedicated machine

Review GodaGoo's always-on architecture docs (link)

Configure Claude Code to run headless (use AI prompt below)

Set up process manager (PM2 or systemd) for auto-restart

Verify Claude Code stays running after closing terminal

Test by sending a request after 30+ minutes of idle

⚠️ GodaGoo's approach: He keeps his laptop running 24/7 for now. For production use, consider a cheap VPS ($5-20/month on DigitalOcean, Hetzner, or AWS Lightsail). But start on your laptop to validate everything works before adding VPS complexity.

🤖 AI Prompt: Configure Headless Claude Code

I need to run Claude Code 24/7 in headless mode on my [Mac/Windows/Linux].

Requirements:
1. Claude Code runs as background process
2. Survives terminal close
3. Auto-restarts on crash
4. Logs output to file for debugging
5. Can receive external triggers (HTTP/webhook)

My OS: [YOUR OS]
My Claude Code location: [PATH]

Generate:
1. Complete setup script for my OS
2. Process manager config (PM2 or systemd)
3. Startup script that runs on boot
4. Health check command I can run to verify it's alive
5. Log rotation config so logs don't fill disk
6. Troubleshooting guide for common issues

Make it copy-paste ready for my OS.

🔧 Technical: 2-Hour Time Limit Strategy

GodaGoo implemented a 2-hour limit on continuous autonomous action — the assistant can work on tasks for up to 2 hours, then must report back. This prevents runaway operations and keeps you in control. Configure this in your Claude Code settings or as a skill that wraps task execution with a timer and mandatory checkpoint.

💻 Cross-Platform Daemon Setup (macOS / Linux / Windows)

macOS — LaunchAgent

Keeps the bot running and restarts on crash. Copy template from repo's daemon/launchagent.plist to ~/Library/LaunchAgents/, edit paths, then:

launchctl load ~/Library/LaunchAgents/com.claude.telegram-relay.plist
launchctl list | grep claude
tail -f ~/Library/Logs/claude-telegram-relay.log

Linux — systemd

Copy daemon/claude-relay.service to /etc/systemd/system/, edit paths and user, then:

sudo systemctl daemon-reload
sudo systemctl enable claude-relay
sudo systemctl start claude-relay
journalctl -u claude-relay -f

Windows — PM2 (Recommended)

npm install -g pm2
pm2 start src/relay.ts --interpreter bun --name claude-relay
pm2 save
pm2 startup   # run the command it outputs

Alternatives: Windows Task Scheduler (built-in) or NSSM (turns any script into a Windows service). See the GitHub repo for full instructions on all three.

🧩 Core Pattern: The Minimal Spawn

The entire relay is built on one simple pattern — spawn the Claude CLI with a prompt and return the output:

claude -p "[prompt]" --output-format text

Why CLI spawn vs. API? The CLI gives you everything: tools, MCP servers, context management, permissions. The API is just the model. Overhead is ~1-2 seconds per spawn — fine for a personal assistant.

For session continuity across messages, use the --resume flag:

claude -p "[prompt]" --resume [sessionId] --output-format text

Mark Step 1 Complete

↑ Return to Top

2

Connect Telegram Bot Relay

CORE

Telegram becomes your universal interface to Claude Code. Text, voice messages, images, files — all processed by your AI. Claude can respond with text, voice replies, files, and images back to you. The BUN relay acts as the bridge between Telegram and your headless Claude Code instance.

❌ Before

• Must be at terminal to use Claude Code
• No mobile access to your AI
• Can't send voice or files remotely
• Zero integration with messaging

✅ After

• Message your AI from anywhere (phone, tablet, any device)
• Send text, voice, images, files — all processed
• AI sends back text, voice replies, files, images
• Full Claude Code power in your pocket

📋 RECIPE

Create a Telegram bot via @BotFather (get your bot token)

Clone GodaGoo's relay repo: claude-telegram-relay

Install dependencies (BUN runtime + Grammy framework)

Configure relay with your bot token + Claude Code endpoint

Start the relay service alongside Claude Code

Test: Send text message → verify response

Test: Send voice message → verify transcription + response

Test: Send image/file → verify processing

💡 Pro Tip: GodaGoo's community already has experience building AI bots in Telegram (their "Jarvis Jr." project). The relay pattern is proven — Telegram as universal interface, Claude Code as brain. All your existing Claude Code skills, MCP servers, and integrations work through Telegram automatically.

🤖 AI Prompt: Set Up Telegram Relay

I need to connect Claude Code to Telegram so I can message my AI from anywhere.

Architecture:
- Telegram Bot ↔ BUN Relay ↔ Claude Code (headless)
- Bidirectional: I send messages, AI responds
- Supports: text, voice messages, images, files
- Grammy framework for Telegram API

I've already:
- Created Telegram bot via @BotFather (token: [REDACT BEFORE PASTING])
- Cloned https://github.com/godagoo/claude-telegram-relay
- Have BUN installed

Help me:
1. Configure the relay with my bot token
2. Connect it to my headless Claude Code instance
3. Set up message routing (text → text, voice → transcribe → process → voice reply)
4. Add file handling (receive files, process, send back results)
5. Security: Only allow MY Telegram user ID to interact
6. Start both services together (Claude Code + relay)

Generate complete config files and startup script.
My OS: [YOUR OS]
My Telegram user ID: [YOUR ID]

🎯 Validation: What "Working" Looks Like

Test These Inputs

• Send "What time is it?" → get text response
• Send voice message → get text summary back
• Send screenshot → get description
• Send PDF → get summary
• Ask it to "check my email" → triggers Gmail skill

Verify These Behaviors

• Response time under 30 seconds
• All Claude Code skills accessible
• MCP servers reachable through Telegram
• Only YOUR messages get processed
• Errors handled gracefully (no crashes)

Mark Step 2 Complete

↑ Return to Top

3

Build Persistent Memory System

INTELLIGENCE

Without memory, your assistant forgets everything between sessions. GodaGoo built a semantic memory system using Supabase that stores conversations, goals, facts, and learnings. When context is needed, the system fetches relevant memories using timestamps, keywords, and semantic search — so the AI knows what you discussed yesterday, last week, or last month.

❌ Before

• Every conversation starts from zero
• Can't recall yesterday's discussion
• No goal tracking across sessions
• Must re-explain context every time

✅ After

• "Remember what we talked about yesterday?" → Yes.
• Semantic search across all past conversations
• Goal detection: facts vs. goals vs. reminders
• Post-call summaries stored automatically

📋 RECIPE

Create free Supabase project (supabase.com)

Set up memory tables: conversations, goals, facts, learnings

Enable pgvector extension for semantic search

Build context fetching logic (timestamps + keywords + semantic)

Integrate memory writes after every conversation

Add goal detection (AI classifies: fact / goal / reminder)

Test: Have a conversation, wait 1 hour, ask "What did we discuss?"

Integrate Claude Code logs/learnings into memory pipeline

🔮 Memory Architecture: GodaGoo's system collects context from three sources: (1) Telegram message history, (2) Semantic memory in Supabase, and (3) Claude Code logs and learnings. When the AI needs context, it pulls from all three — checking timestamps, keywords, and semantic similarity. During conversations, it detects goals vs. facts vs. reminders and categorizes automatically.

🤖 AI Prompt: Build Memory System

Build a persistent memory system for my 24/7 AI assistant using Supabase.

REQUIREMENTS:
1. STORAGE:
   - conversations table (timestamp, participants, summary, full_text)
   - goals table (description, status, created_at, deadline)
   - facts table (key, value, source, confidence)
   - learnings table (insight, context, created_at)

2. SEMANTIC SEARCH:
   - Enable pgvector in Supabase
   - Generate embeddings for all stored content
   - Similarity search with configurable threshold
   - Combined search: timestamp + keywords + semantic

3. CONTEXT FETCHING:
   - fetchRecentContext(hours=24) → last day's conversations
   - fetchByKeyword(terms) → keyword-matched memories
   - fetchSemantic(query, limit=5) → most relevant memories
   - fetchGoals(status='active') → current goals

4. AUTO-CLASSIFICATION:
   - During conversation, detect and categorize:
     * Goals ("I want to...", "by next week...")
     * Facts ("My dog's name is...", "The API key is...")
     * Reminders ("Don't forget to...", "Remind me...")
     * Learnings ("Interesting, so...")

5. WRITE PIPELINE:
   - After every conversation: store summary + full text
   - After every voice call: store transcript + summary
   - Periodically: consolidate + deduplicate

Generate:
- Supabase SQL schema (tables, indexes, functions)
- Node.js/Python integration code
- Context fetching utility functions
- Auto-classification prompt for Claude
- Test script to verify the system works

Mark Step 3 Complete

↑ Return to Top

4

Add Voice Calls (Bidirectional)

OPTIONAL

This is the wow factor — call your AI assistant on the phone and have a natural conversation. GodaGoo uses 11 Labs for voice synthesis, Twilio for the phone number, and a context pipeline that loads memory before each call. After the call ends, the full transcript gets stored in memory automatically. You can trigger the call, or the AI can call you.

❌ Before

• Text-only interaction
• Can't talk to your AI on the go
• No hands-free operation
• No post-call action execution

✅ After

• Call your AI from any phone
• AI can call YOU proactively
• Natural voice conversation with full context
• "Go research X, create analysis, call me back with results"

📋 RECIPE

Create 11 Labs account and set up conversational voice agent

Create Twilio account and provision a phone number

Build context injection pipeline (loads memory before call starts)

Connect 11 Labs voice agent to Claude Code via API

Add caller ID security (only YOUR number gets through)

Build post-call pipeline: transcript → summary → Telegram → memory

Test: Call your AI → have conversation → verify transcript stored

Test: Trigger AI to call YOU → verify it rings and converses

⚠️ Security Critical: GodaGoo emphasizes this — if you expose the phone number, ANYONE can call and access your memory, tools, and controls. Implement caller ID verification. GodaGoo built security measures so unknown callers can't get through the system.

💡 Post-Call Actions are the real power: GodaGoo doesn't just talk — he gives commands during calls. "Go research this topic. Find this PDF. Save it to Google Drive. Create full analysis. Evaluate if it's worth making a video on. Package it with titles. Write a script. Call me back." The AI executes ALL of that after the call ends.

🤖 AI Prompt: Set Up Voice Calling

I want bidirectional voice calling with my 24/7 AI assistant.

STACK:
- 11 Labs: Conversational voice agent (speech synthesis)
- Twilio: Phone number provisioning + telephony
- Claude Code: Brain (processes requests, executes actions)
- Supabase: Memory (context loaded before call, transcript stored after)

BUILD:
1. INBOUND CALLS (I call AI):
   - Twilio receives call on my number
   - Verify caller ID matches my phone number
   - If unauthorized → reject or play generic message
   - If authorized → load context from memory system
   - Connect to 11 Labs conversational agent
   - 11 Labs agent powered by Claude Code

2. OUTBOUND CALLS (AI calls me):
   - Triggered by: proactive check-in, task completion, urgent alert
   - Claude Code triggers Twilio to call my number
   - Same 11 Labs agent with pre-loaded context

3. CONTEXT PIPELINE:
   - On call start: fetch last 24h conversations + active goals + pending tasks
   - During call: real-time processing of my requests
   - On call end: capture full transcript

4. POST-CALL ACTIONS:
   - Store transcript in Supabase memory
   - Generate summary → send to Telegram
   - Execute any action items mentioned during call
   - Update goals/tasks based on conversation

5. SECURITY:
   - Caller ID whitelist (my numbers only)
   - Rate limiting on calls
   - Auto-hangup after 30 min max

Generate complete integration code and config for each component.

Mark Step 4 Complete

↑ Return to Top

5

Configure Proactive Check-Ins

GAME-CHANGER

This is the feature everyone went crazy about with ClawdBot — your AI reaches out to YOU. Every 30 minutes, it checks your calendar, email, projects, tasks, and partnerships. But there's a critical framework: it must decide whether to skip, text, or call. Without this filter, you get noise. With it, you get a proactive assistant that only interrupts when it matters.

❌ Before

• You must remember to check everything
• Important emails slip through
• Meetings sneak up without prep
• AI only works when you ask it to

✅ After

• AI monitors calendar, email, tasks on schedule
• Smart filtering: skip / text / call framework
• "New sponsorship offer" → "Should I run evaluation?"
• Cross-references email with partnerships in Notion

📋 RECIPE

Define your check-in sources (Gmail, Google Calendar, Notion, etc.)

Set up cron/scheduler: check every 30 minutes

Build the SKIP / TEXT / CALL decision framework (prompt below)

Add last-check-in context (prevents repeating same alerts)

Integrate goal tracking: detect goals from conversations automatically

Add partnership/contact cross-referencing (email sender → Notion lookup)

Test: Let it run for 4+ hours, review what it reports vs. skips

Tune the framework based on signal-to-noise ratio

🔮 Critical Insight from GodaGoo: Without a decision framework, every check-in sends "Hey, you got an email" — pure noise. The AI needs context about what it said LAST time, what's actually important, and what warrants interruption. It keeps a log of what it's already pinged you about. This is the difference between a useful proactive assistant and an annoying notification bot.

⚠️ Smart Gating Hard Rules (non-negotiable): GodaGoo built hard boundaries that Claude cannot override:

• No contact if checked in <2 hours ago — unless truly urgent
• 7-10am is sacred creative time — zero interruptions during peak hours
• Quiet after 10pm — everything can wait until morning
• Knows pickup times — reminds 30min before, not during

💡 Three intelligence features that prevent spam:
Reply Awareness: Checks your sent folder. If you already replied, it flags "ALREADY REPLIED on Jan 21" and won't nag you about it.
Partnership Memory: Tracks contact history. When you get the 4th email from the same company, it shows: "4th contact attempt. You quoted $8K on Jan 20. Status: awaiting response."
Action Buttons: Telegram messages include inline buttons — [📊 Evaluate] [😴 Snooze] for emails, [✅ Yes, call] [❌ Not now] for call escalation. One tap to act.

🤖 AI Prompt: Build Proactive Check-In System

Build a proactive check-in system for my 24/7 AI assistant.

SCHEDULE: Every 30 minutes via cron/launchd

DATA COLLECTION (all in parallel):
1. Gmail: Unread from last 7 days, identify partner/sponsor, check if already replied
2. Google Calendar: Today's events + next 3 days preview + conflict detection
3. Notion Tasks: Tasks marked "Now" + due today + P1 priority this week
4. Partnerships: Notion history lookup (previous quotes, outreach count)
5. Memory: Current goals, key facts, pending items
6. History: Last 3 days of chat + previous check-ins + call history

HARD GATING RULES (non-negotiable):
- No contact if checked in less than 2 hours ago (unless truly urgent)
- 7-10am is sacred creative time — ZERO interruptions
- Quiet after 10pm — everything waits until morning
- Know pickup/family times — remind 30min before, not during

DECISION FRAMEWORK (Critical — prevents noise):
For each finding, classify as:

SKIP (don't notify):
- Newsletter/marketing emails
- Already-reported items (check last check-in log)
- Low-priority task updates
- Calendar events 2+ hours away with no prep needed

TEXT via Telegram:
- New email from active partnership/client
- Calendar event in <1 hour needing prep
- Task deadline approaching (24 hours)
- Goal milestone reached
- Offer: "Want me to run [specific action]?"

CALL (urgent):
- Email from VIP contact requiring immediate response
- Calendar conflict detected
- Critical deadline missed
- System error requiring attention

ANTI-REPEAT LOGIC:
- Store each check-in's findings in memory
- Before sending notification, verify it wasn't already reported
- Track "conversation state" per topic (new → reported → acknowledged → resolved)

CONTEXT ENRICHMENT:
- When email arrives from known contact:
  * Check Notion for active projects with that person
  * Pull last 3 interactions from memory
  * Show contact attempt count: "4th email from Agency X"
  * Include previous quotes/pricing if applicable
  * Include relevant context in notification
- REPLY AWARENESS: Check sent folder before alerting
  * If already replied, flag "ALREADY REPLIED on [date]" and don't nag
- TELEGRAM ACTION BUTTONS: Include inline buttons with each notification
  * Emails: [📊 Evaluate] [😴 Snooze] [✉️ Reply]
  * Meetings: [📝 Prep] [⏰ Remind Later]
  * Calls: [✅ Yes, call] [❌ Not now]

Generate:
1. Cron job configuration
2. Check-in orchestration script
3. Decision framework prompt (for Claude to evaluate each finding)
4. Anti-repeat logic with check-in log
5. Notification templates (Telegram + voice call trigger)
6. Test mode (run once and show what WOULD be sent)

📊 GodaGoo's Check-In Architecture

What Gets Checked

• Calendar events
• Email inbox
• Active projects
• Task lists
• Partnership status
• Goal progress

Smart Filter Examples

• New sponsorship email → cross-ref with Notion partnerships → "New inquiry — run evaluation?"
• Same email reported last check → SKIP
• Meeting in 45 min → "Prep needed?"
• Newsletter → auto-SKIP

Mark Step 5 Complete

↑ Return to Top

6

Build Observability Dashboard

CONTROL

GodaGoo built a system observatory to see if Telegram bot is online, if Supabase is connected, uptime tracking, and goal monitoring. ClawdBot users often don't know what their agent is doing — you will. This is your control panel for the entire system.

❌ Before

• No idea if assistant is running
• Can't see what it's doing autonomously
• Errors happen silently
• No uptime or health tracking

✅ After

• Dashboard shows all service statuses
• Live feed of agent actions
• Uptime tracking (hours, days)
• Goal tracking visualization

📋 RECIPE

Define service health checks (Telegram bot, Supabase, Claude Code, voice)

Build simple HTML dashboard (Claude Code can generate this)

Add live action feed (last 20 agent actions with timestamps)

Add uptime counter and service status indicators

Add goal tracking panel (pulls from Supabase)

Test: Force-kill a service and verify dashboard shows it offline

⚠️ Observability is non-negotiable: GodaGoo specifically calls this out — ClawdBot users don't know what their agent is doing. You MUST have visibility into autonomous actions, especially with a 2-hour execution limit. If the AI goes off-rails, you need to see it immediately.

🤖 AI Prompt: Generate Observatory Dashboard

Build an observability dashboard for my 24/7 AI assistant system.

SERVICES TO MONITOR:
1. Claude Code (headless) - is process running?
2. Telegram Bot Relay - is bot online and responding?
3. Supabase - is database connected?
4. Voice System (11 Labs + Twilio) - are services reachable?
5. Proactive Check-Ins - is cron running? Last execution time?

DASHBOARD FEATURES:
- Service status cards (green/red with uptime counter)
- Live action feed (last 20 agent actions with timestamps)
- Goal tracker (active goals from Supabase with progress)
- Memory stats (total memories, recent additions)
- Error log (last 10 errors with details)
- Cost tracker (if applicable)

TECH:
- Simple HTML/CSS/JS (no framework needed)
- Polls API endpoints every 30 seconds
- Runs on localhost:3000
- Mobile responsive

Generate:
1. Complete HTML dashboard file
2. Backend API endpoints (Node.js/Express)
3. Health check functions for each service
4. Auto-refresh with visual indicators
5. Alert system (Telegram notification if service goes down)

Mark Step 6 Complete

↑ Return to Top

✓

Validate Your Complete Setup

PROOF

Run through this final checklist to verify your entire 24/7 AI assistant is operational. Every feature should be independently testable.

🔮 GodaGoo's Timeline: Built on Sunday, fine-tuned Monday, using all day Tuesday, filming the video Wednesday. If you already have Claude Code with skills, the core build is genuinely ~2 hours. The fine-tuning and proactive check-in calibration takes another 1-2 days of real usage.

🤖

AI Prompts Library

REFERENCE

All prompts from each step collected here for easy reference and batch copying.

Prompt 1: Assess Your Starting Point (Step 0)

I want to build a 24/7 AI assistant accessible via Telegram with:
1. Always-on Claude Code (headless)
2. Telegram bot relay for messaging
3. Persistent memory (Supabase or local)
4. Bidirectional voice calling (11 Labs + Twilio)
5. Proactive check-ins (calendar, email, tasks)
6. Observability dashboard

Audit my current Claude Code setup and create a GAP ANALYSIS:
- What I already have (✅) vs. What I need to build (❌)
- Estimated time per gap
- Recommended build order
Output as a structured checklist.

Prompt 2: Configure Headless Claude Code (Step 1)

Run Claude Code 24/7 headless on [YOUR OS].
Requirements: background process, survives terminal close, auto-restart on crash, logs to file, accepts external triggers.
Generate: setup script, PM2/systemd config, startup-on-boot script, health check command, log rotation, troubleshooting guide. Copy-paste ready.

Prompt 3: Telegram Bot Relay (Step 2)

Connect Claude Code to Telegram via BUN relay + Grammy.
My bot token: [TOKEN]. My Telegram user ID: [ID].
Set up: message routing (text/voice/files), security (only my ID), bidirectional file transfer.
Generate complete config + startup script for [YOUR OS].

Prompt 4: Memory System (Step 3)

Build persistent memory using Supabase with pgvector.
Tables: conversations, goals, facts, learnings.
Features: semantic search, auto-classification (goal/fact/reminder), anti-duplicate, context fetching by time+keyword+semantic.
Generate: SQL schema, integration code, classification prompt, test script.

Prompt 5: Voice Calls (Step 4)

Build bidirectional voice calling: 11 Labs (voice) + Twilio (phone) + Claude Code (brain) + Supabase (memory).
Inbound: verify caller ID → load context → connect to voice agent.
Outbound: AI triggers call to my number with pre-loaded context.
Post-call: transcript → summary → Telegram → memory.
Security: caller ID whitelist, rate limiting, 30 min max.
Generate complete integration code.

Prompt 6: Proactive Check-Ins (Step 5)

Build proactive check-in system running every 30 min.
Sources: Gmail, Calendar, Notion, partnerships.
Decision framework: SKIP (noise) / TEXT (important) / CALL (urgent).
Anti-repeat: log last check-in findings, don't re-report.
Context enrichment: cross-reference senders with active projects.
Generate: cron config, orchestration script, decision prompt, anti-repeat logic, notification templates.

Prompt 7: Observability Dashboard (Step 6)

Build HTML observability dashboard for my AI assistant.
Monitor: Claude Code, Telegram Bot, Supabase, Voice System, Check-Ins.
Features: service status cards, live action feed, goal tracker, error log, auto-refresh every 30s.
Generate: complete HTML file + backend API + health checks + Telegram alert if service goes down.

📚

Resources & Tools

Build Your Own 24/7 AI Assistant with Claude Code

⚠️ PREREQUISITES

ClawdBot (Hosted)

Your Own (Claude Code)

macOS — LaunchAgent

Linux — systemd

Windows — PM2 (Recommended)

Test These Inputs

Verify These Behaviors

What Gets Checked

Smart Filter Examples

Essential Resources

Tools & Services

Want to Build AI Systems That Run 24/7?