Claude Code study: knowing the problem beats knowing only code
Anthropic analyzed about 400,000 Claude Code sessions from about 235,000 people between October 2025 and April 2026, using a privacy-preserving method. In a typical session, people made about 70% of the planning decisions, such as what to build and what counts as done, while Claude made about 80% of the execution decisions, such as which files to edit and which commands to run. About 56% of sessions involved writing new code, fixing code, testing, or automation, and 17% involved operating software such as deploying, configuring, or running systems.
Over seven months, debugging fell from 33% of sessions to 19%, while operating software, , and writing documents became more common. More expert users got Claude to do more work from each instruction. Novice sessions led to about 5 Claude actions and 600 words of output per prompt, while expert sessions led to about 12 actions and 3,200 words.
Success also rose with : novice sessions reached 15% of the time, while intermediate or higher sessions reached 28% to 33%. In sessions that produced code, software workers and non-software workers had almost the same partial success rate, 89% and 88%, which suggests that understanding the problem can matter more than having a traditional coding job.
Key points
- The study covered about 400,000 Claude Code sessions from October 2025 to April 2026.
- People mostly chose what to do, while Claude mostly chose how to carry it out.
- Debugging fell from 33% to 19% of sessions, while operations, , and document work grew.
- Expert users got longer action chains and more output from each prompt.
- Non-software workers nearly matched software workers on partial success when sessions produced code.