Descriptions:
Peter Yang sits down with Alex Albert — Anthropic’s first prompt engineer, former Head of Developer Relations, and now Product Manager on the research team — for an unusually candid look at how Claude models are conceived, built, and refined. Albert explains that research PMs attach to a model from its earliest ideation phase, speccing out capability requirements (coding, knowledge work, spreadsheet tasks) and tracking the model through training all the way to launch, with ongoing loops of customer interviews and internal feedback to inform the next generation.
Several disclosures stand out. Albert describes a concept he calls ‘dreaming’: when an agent is idle or waiting between tasks, it actively reviews its memory store, identifies contradictions, and prunes stale entries — a design choice that treats memory hygiene as an ongoing background process rather than a one-time setup. He also explains why an AI agent’s character and values become load-bearing properties as task length and autonomy increase, and why Anthropic treats the model itself as a product with a spec rather than purely a research artifact.
On the practical side, Albert shares how he uses Claude Code with access to internal databases and Slack to compress what used to be multi-day data investigations into ten-minute sessions, and why Anthropic’s Co-work desktop product has become his primary interface for document review and strategic brainstorming. The conversation offers a rare first-person account of what building at the frontier of AI development actually looks like from the inside.
📺 Source: Peter Yang · Published May 17, 2026
🏷️ Format: Interview
