Anthropic studied 400,000 Claude Code sessions and found the thing you bring isn't code

This is the most useful piece of research I've read all month, and it quietly demolishes a belief a lot of people hold. Anthropic analyzed roughly 400,000 Claude Code sessions across 235,000 people from October 2025 to April 2026, and the headline is that expert users hit 33% verified task success versus 28% for intermediate and 15% for novice. More than 2x from bottom to top. (Anthropic)

But the interesting part is what "expert" means. It's not "writes better code." Experts trigger about 12 Claude actions per prompt and write ~3,200 words, against 5 actions and ~600 words for novices. They're doing more steering, more decomposition, more context-setting. And the line that stopped me cold: nearly every occupation succeeds at coding tasks at roughly the same rate as software engineers. A domain expert who can't write a for-loop wins about as often as a senior engineer, as long as they understand the problem deeply.

I've felt this in my own work without naming it. When I shipped Document Domain Agents, the slow part was never the LangGraph wiring. It was knowing what a "good" GraphRAG retrieval looked like for legal documents, which is domain understanding, not coding. The model handled the "how." I owned the "what." Anthropic just put 235,000 people behind that intuition.

The actionable read for builders is uncomfortable if you've been telling yourself your edge is typing speed or syntax fluency. It isn't. The scarce input is understanding the problem well enough to specify it, verify it, and catch the model when it confidently goes sideways. That maps almost exactly onto the context-engineering story below: the people winning aren't the ones with the cleverest prompts, they're the ones who front-load the most useful context and review the plan, not the diff.

So stop optimizing the wrong muscle. If you want to get better at agentic coding, get deeper in a domain. Learn what correct looks like in a field where you can verify it. The model will close the syntax gap. It can't close the taste gap, and it can't tell you what's worth building.