Claude Code: Experience Still Matters in AI

Anthropic analyzed around 400,000 Claude Code sessions and found a clear pattern: AI can handle much of the execution, but users get better results when they understand the problem they are trying to solve.

The study, based on sessions from around 235,000 people between October 2025 and April 2026, examines how Claude Code is used, which decisions each side makes and when the work succeeds. The main conclusion is straightforward: experience in a specific area seems to matter more than knowing how to code.

People decide what to do, and AI decides how to do it

In a typical session, the user makes about 70% of the planning decisions: what to build, which approach to take and when to consider the work finished. Claude, by contrast, makes around 80% of the execution decisions: which files to modify, what code to write and which commands to run.

The division of labor is quite concrete. You might tell Claude to prepare a tool that detects missing clauses in contracts. The agent decides how to read the documents and generate the code, but it needs you to know which legal rules to apply.

Each user instruction triggers an average chain of about 10 Claude actions. The tool reads files, modifies code, runs commands and returns nearly 2,400 words per turn. A typical session has around four turns, although some are far more autonomous.

Experience helps you get more from the tool

Anthropic classified the level of experience each user showed in the specific task, from beginner to expert. This is not necessarily the same as their professional title. An engineer may be a beginner when working with Rust for the first time, while an accountant may demonstrate expertise by knowing exactly which rules a financial script must follow.

The difference shows up in how much work each instruction activates:

• In beginner sessions, each instruction triggers about 5 actions and roughly 600 words of response.
• In expert sessions, it triggers about 12 actions and around 3,200 words.

The explanation is not that experts write better instructions because they know more code. Their advantage is that they define the goal more clearly, spot errors and know what to ask Claude to check.

Experts also finish more sessions successfully

The study uses two main measures. One estimates whether the user achieved what they wanted, while the other requires verifiable signals, such as passed tests, confirmed Git changes or an explicit indication that the result works.

Under the stricter measure, only 15% of beginner sessions achieve verified success. In sessions classified as intermediate or expert, the figure rises to between 28% and 33%. If partial success is included, the percentages rise from 77% among beginners to 91% or 92% in the other groups.

The difference becomes more pronounced when something goes wrong. Among sessions with problems, verified success rises from 4% among beginners to 15% among experts. In addition, 19% of beginner sessions that run into trouble are abandoned, compared with between 5% and 7% of the rest.

The biggest improvement comes when moving from beginner to intermediate level. Mastery adds something, but the jump is smaller. Having a functional understanding of the field appears to capture much of the benefit.

Claude Code is already being used beyond programming

Between October 2025 and April 2026, the type of work done with the tool changed. Sessions devoted to fixing code fell from 33% to 19%. At the same time, tasks surrounding code grew:

• Operating software rose from 14% to 21%.
• Writing and analyzing data increased from approximately 10% to 20%.
• Planning and exploration accounted for 14% of sessions.

Overall, around 56% of sessions involved creating, repairing, testing or orchestrating code. The rest focused on operating systems, understanding projects, analyzing data or producing documents and presentations.

Anthropic also estimated the value of the tasks by comparing them with jobs posted on freelance platforms. According to that reference, the value of a typical session increased 27% during the period analyzed. The company warns that this is not a literal amount of money, but an approximate way to compare the complexity and relative value of the tasks.

Knowing how to code is not the only advantage

In sessions that produced code, users in software-related professions achieved verified success of 34%, compared with 29% among people working in other fields. For total or partial success, the difference was much smaller: 89% versus 88%.

The ten main occupations in the dataset were all within seven points of software professionals. They included fields such as management, finance, sales, law and science.

For you, the implication is direct: a tool like Claude Code can let you build an automation, analyze data or launch a system without mastering every technical detail. But it does not remove the need to know what result you need, which constraints exist and how to check whether the answer is correct.

The study has important limitations. Anthropic could not verify whether the code was actually used afterward, whether it generated economic value or whether the user was satisfied outside the session. In addition, many of the conclusions came from classifiers that interpret conversations, not from human reviews of every case.

What to watch now is whether this experience advantage shrinks as models improve. If it does, that would mean AI is starting to contribute more judgment about the problem on its own. If it does not, the market may increasingly value the ability to understand a domain and direct agents that execute the work, while placing less value on the ability to write code.