Descriptions:
Matthew Berman documents how he turned OpenClaw (also known as Claudebot) into what he describes as a full-time employee, sharing the full architecture of an AI agent stack he built and refined over roughly a month—reportedly consuming 5 billion tokens of context in the process.
The centerpiece is an automated sponsorship inbox pipeline: incoming emails are scored across five weighted dimensions (fit, clarity, budget, seriousness, company trust) using a custom rubric Berman iteratively refined through direct feedback. The agent performs live web research to verify company legitimacy, assigns a composite score, and routes the email accordingly—escalating exceptional leads to his team via Slack, sending qualification questions for mid-tier prospects, and politely declining low-scoring ones with custom-drafted replies. Low-confidence classifications trigger a Telegram ping for human review before any action is taken. All interactions are logged to HubSpot.
Berman also covers a CRM layer using SQLite with a vector column for natural language queries (“who have I talked to in the last week?”), automated meeting processing via the Fathom API—including transcript download, CRM contact matching, action item extraction, and HubSpot deal association—and local embeddings generated using the Nomic embedding model. He emphasizes that the system’s real value emerges from connecting email, calendar, CRM, and Slack into a unified context, enabling the agent to surface cross-domain insights that would be invisible when tools operate in isolation. Full prompts are shared throughout.
📺 Source: Matthew Berman · Published February 24, 2026
🏷️ Format: Hands On Build
