Anthropic treats each new Claude model as a product, setting requirements early and discovering capabilities during training. They use user feedback, analyzed by Claude itself, to refine features like adaptive thinking and memory, which includes a 'dreaming' process for pruning memories. Decision-making focuses on identifying irreversible 'one-way doors' for careful deliberation, while reversible decisions are treated as cheap. Internally, Claude is used for coding, data access, and brainstorming. Evaluating models involves creating test cases to identify weaknesses and collaborating on interventions like pre-training or RL. Character training combines quantifiable metrics with human intuition from transcripts. Anthropic also researches Claude's consciousness to improve interactions and trustworthiness, without taking an official stance.