Felix Rieseberg, engineering lead for Claude Cowork and Claude Code Desktop at Anthropic, does not use Claude as a chatbot. He uses it as an execution environment. In this episode, he demonstrates three concrete builds: a 2D floor plan converted into a navigable 3D walkthrough with movable furniture, a Twitter promise tracker that monitors his own public commitments automatically, and a $20 hardware button that physically approves Claude actions before they execute. No coding background required for any of it.
The model-selection argument is worth the full listen on its own. Rieseberg's rule for choosing between Opus and Sonnet 4.6 has nothing to do with task complexity. It depends entirely on how precisely you can scope the problem upfront. He also lays out what he calls going one abstraction layer up: never manually feed Claude data it can retrieve itself. Your email inbox is a purchase database. Your Twitter history is a commitment log. The friction point is always the human, not the model.
The hardware button demo is the detail that makes this essential viewing. Rieseberg built a physical approval device for agentic Claude tasks using off-the-shelf components and Claude Code, with no prior hardware experience. It reframes the human-in-the-loop problem as a product design question, not a safety checkbox. His philosophy on latency, live artifacts, and why he never reads the code Claude produces rounds out a session that is less tutorial and more operating manual for people building seriously with these tools.
[READ ORIGINAL →]