Matt Berman OpenClaw Tips — Analysis¶
Source: "I've spent 5 BILLION tokens perfecting OpenClaw" — YouTube Analyzed: 2026-03-18
Summary¶
Matt spent 200+ hours and billions of tokens optimizing his OpenClaw setup. 34-minute video covering threading, models, sub-agents, security, crons, documentation, and notification batching.
Tip-by-Tip Analysis: Do We Have It?¶
1. Telegram Threads ⭐ HIGH PRIORITY — WE SHOULD IMPLEMENT¶
What Matt does: Creates a Telegram group with just himself + bot, then uses topics/threads to separate conversations by subject (CRM, knowledge base, cron updates, etc.). Each topic gets its own session and context window.
Our status: ❌ We're running everything in one flat Telegram DM. Jeff specifically mentioned interest in this.
Why it matters: This is probably the single biggest unlock for us. Each thread gets its own lean context window. No cross-topic contamination. Jeff can jump between "BioThread research" and "MC bug fixes" without telling me to "hold that thought." This directly addresses our context bloat problem AND our token spend.
Implementation: Create a Telegram group, add Jules, create topics for: General, APA Initiative, BioThread, MC Development, Ideas/Brainstorm, Operations/Crons.
2. Voice Memos ✅ ALREADY AVAILABLE¶
What Matt does: Uses Telegram's built-in voice memo feature to talk to OpenClaw hands-free.
Our status: ✅ Available natively in Telegram. Jeff just needs to use the mic button.
3. Multi-Model Strategy ✅ ALREADY DOING THIS WELL¶
What Matt does: Different models for different tasks. Opus for main chat, Sonnet for cheaper tasks, GPT 5.4 as fallback, local Qwen3.5, Nomic for embeddings.
Our status: ✅ We have this nailed. Documented in MEMORY.md as "Local Model Strategy." Opus for main, Haiku for crons, local GLM for heartbeats, Nomic for embeddings, Sonnet for sub-agents.
4. Per-Thread Model Assignment 🔄 IMPLEMENT WITH THREADS¶
What Matt does: Assigns different models to different threads. Q&A thread gets a cheaper model, coding thread gets Opus.
Our status: ❌ Requires threads first. Once we have threads, this is a quick win.
5. Fine-Tuning Local Models for Specific Tasks 📋 BACKLOG¶
What Matt does: Collected email labeling data from Opus, fine-tuned Qwen3.5 9B to replace it. Free inference.
Our status: ❌ Not doing this yet. Interesting for future — could fine-tune local models for repetitive tasks like SigInt triage, idea evaluation, etc.
6. Sub-Agent Delegation ✅ ALREADY DOING¶
What Matt does: Delegates all coding to sub-agents, uses cursor/Claude Code, keeps main agent unblocked.
Our status: ✅ We delegate to Melody (Claude Code), Atlas (research), Quinn (QA). Main session stays conversational.
7. Model-Specific Prompt Files 📋 CONSIDER¶
What Matt does: Separate prompt files optimized for each model (Opus vs GPT). Nightly cron syncs them.
Our status: ❌ We use one set of workspace files for all models. This could improve quality when falling back to GPT or Gemini, but adds complexity. Low priority until we use fallbacks more often.
8. Crons Scheduled Overnight ✅ PARTIALLY DOING¶
What Matt does: Runs all non-time-sensitive crons between midnight and 6 AM to avoid quota competition with daytime usage.
Our status: 🔄 Some crons are overnight (backup 11 PM, consolidation 10 PM), but Atlas scan is 7 AM and heartbeat is hourly. Could shift more to overnight.
9. Use Subscription Auth (Not API) ⚠️ ACTION NEEDED¶
What Matt does: Uses Anthropic subscription through the Agents SDK (not raw OAuth) to stay within TOS. Uses OpenAI Codex OAuth for GPT models.
Our status: ⚠️ We're using raw Anthropic OAuth (sk-ant-oat-*), which Matt says is technically not TOS-compliant. He says to use the Agents SDK instead, which Anthropic has explicitly approved for OpenClaw use. We should switch.
Action: Tell OpenClaw to install and use the Anthropic Agents SDK. This keeps us on the subscription (flat rate, no per-token cost) but within TOS.
10. Security Hardening 📋 REVIEW¶
What Matt does: Multi-layer prompt injection defense (text sanitization + frontier model scanner), outbound PII redaction, granular permissions, approval system for destructive actions, runtime governance (spending caps, loop detection).
Our status: 🔄 We have some basics (exec security, approval-gated sending), but no prompt injection scanning, no PII redaction layer, no spending caps. Worth reviewing against his security article.
11. Log Everything ✅ PARTIALLY¶
What Matt does: Logs all system activity, reviews logs every morning for errors.
Our status: ✅ OpenClaw logs to /tmp/openclaw/, we have MC event logging, and session JSONL files. Could be more structured. Morning log review is a good standup addition.
12. Document Everything ✅ DOING WELL¶
What Matt does: PRD, use cases, workspace files, prompting guides, security docs, learnings.md.
Our status: ✅ We have MEMORY.md (learnings), skills, initiative docs, design system, SOUL.md, etc. Our documentation game is solid.
13. Git Backup ✅ ALREADY DOING¶
What Matt does: Version control + GitHub backup.
Our status: ✅ Nightly cron backs up to GitHub (Vivere-Vitalis-LLC/openclaw-backup).
14. Testing 📋 BACKLOG¶
What Matt does: Writes tests for all code.
Our status: ❌ MC has no test suite. Not critical now but important before shipping APA to customers.
15. Notification Batching ⭐ IMPLEMENT¶
What Matt does: Low-priority notifications batched every 3 hours. Medium every hour. Critical = immediate.
Our status: ❌ We've been fighting notification noise all day (heartbeat leaks, work pulse messages). Batching would solve this elegantly.
Implementation: Create a notification batching cron that collects low-priority events and sends a digest every 3 hours.
16. Build OpenClaw Externally ✅ ALREADY DOING¶
What Matt does: Uses Cursor for code, Telegram for chat.
Our status: ✅ We use Claude Code for MC builds, Telegram for conversation. Same pattern.
Priority Implementation List¶
- Telegram Threads — biggest unlock, directly reduces token spend and improves UX
- Agents SDK for Anthropic — TOS compliance, keeps subscription pricing
- Notification Batching — stop the notification noise permanently
- Shift crons to overnight — preserve daytime quota
- Morning log review — add to standup routine
- Security hardening review — audit against Matt's article