Mike Krieger: How Claude writes 90% of Anthropic's code

Krieger describes Claude Opus 4 surfacing novel, creative angles on Anthropic’s product strategy that he hadn’t considered, shifting his view of when models would show independent, non-obvious thinking rather than just safe, generic feedback.

Anthropic had to re-architect its merge queues and rethink code review because AI massively increased pull request volume; the new constraints are alignment, decision-making, deployment pipelines, and ensuring coherence and launch strategy, not raw implementation speed.

The highest leverage at Anthropic comes from PMs working directly with researchers on post-training and fine-tuning (e.g., Claude Skills + Artifacts), shipping experiences that external builders couldn’t replicate with off-the-shelf prompting alone.

Krieger argues product’s enduring edge is deciding where to play and win, making powerful capabilities comprehensible and trusted, and closing the gap between what models can do and how people actually use them, rather than merely “drawing UIs around a giga-brain.”

Instead of defaulting to “ask Claude,” Krieger encourages his children to design experiments and reason for themselves, aiming to cultivate the scientific process and confidence in their own judgments so they don’t outsource all cognition to AI.