Descriptions:
Matthew Berman compiles 14 advanced best practices for OpenClaw — Anthropic’s Claude-based personal AI agent framework — drawn from over 200 hours of hands-on use. The video opens with an Nvidia CEO Jensen Huang clip calling OpenClaw “the number one open source project in the history of humanity,” framing the stakes before Berman digs into practical configuration advice.
The highest-impact recommendation is using Telegram’s thread feature to partition OpenClaw conversations by topic. Each thread maintains its own context window, preventing cross-topic memory contamination — Berman attributes his complete absence of the memory issues commonly reported by other users to this single architectural choice. Other techniques include Telegram voice memos for hands-free task delegation while driving, integrating the Here platform for hosting agent-created web artifacts, and a structured sub-agent delegation pattern where the main agent stays unblocked by farming out discrete tasks to specialized agentic harnesses.
A particularly detailed section addresses multi-model prompt optimization. Because Claude Opus 4.6 performs worse when given all-caps instructions or told what not to do — while GPT-5.4 responds well to both — Berman recommends maintaining separate prompt files per model, generated automatically by having OpenClaw read each lab’s official prompting guidelines and produce an optimized variant. The strategy extends to soul files, memory files, and skills files. The video serves as a practical reference for power users who want to systematically extract more performance from an OpenClaw setup running multiple frontier models in parallel.
📺 Source: Matthew Berman · Published March 18, 2026
🏷️ Format: Workflow Case Study
