What 400,000 Claude Code sessions say about human expertise

What happened

Anthropic analysed about 400,000 Claude Code sessions to understand how expertise changes the way people work with a coding agent. Its researchers found a recurring pattern: skilled users often kept the larger plan in their own hands while delegating well-bounded execution to the model.

Why it matters

The popular version of prompt advice focuses on wording. This research points to something deeper: good results depend on task design. Experienced developers create useful checkpoints, provide the right context and notice when an apparently plausible answer does not fit the system around it.

That makes AI fluency less like learning a magic sentence and more like learning to manage a fast, capable colleague who lacks the full history of the work.

What is confirmed

The study describes behaviour observed in Anthropic’s own coding product. It suggests that human planning remains important even when a model performs much of the implementation. The paper does not show that every expert works the same way or that the pattern transfers unchanged to law, medicine or design.

A practical takeaway

Define the outcome, expose the relevant constraints and ask for work in reviewable pieces. Let the agent move quickly inside those boundaries. Keep architecture, trade-offs and final accountability with a person who understands the wider system.