You Don't Need to Code to Ship With AI — Here's the Data

You don't need to code to ship with AI — and a 400,000-session study from Anthropic is the strongest evidence yet. In June 2026 the company analyzed roughly 400,000 Claude Code sessions from about 235,000 users and found that people in non-software occupations reached verified success at nearly the same rate as software engineers. The thing that actually separated good sessions from wasted ones was not coding ability. It was domain knowledge — knowing precisely what a correct result looks like before the agent starts.

That is the encouraging headline. The useful part is what comes after it: the study quietly hands you a self-diagnostic. Its classifier rated each session by reading three behavioral signals, and those same three signals double as a checklist you can run on your own prompts. Get them right and you behave like an expert user regardless of whether you have ever opened a terminal before.

This is the practical playbook. We cover what the data actually says (and where it's noisy), the three-signal self-test, what an expert session looks like next to a novice one, why aiming for "intermediate" captures most of the gain, and a five-step routine you can apply to your next AI build. If you want the strategic argument for why this shift creates a durable advantage, read its companion — our deeper analysis on why domain expertise is the new moat in agentic coding. This post is the how-to.

On June 16, 2026, Anthropic published "Agentic Coding and Persistent Returns to Expertise," a privacy-preserving analysis of roughly 400,000 Claude Code sessions from about 235,000 users between October 2025 and April 2026. Rather than read transcripts by hand, the researchers built a classifier on Claude Sonnet 4.6 that rated each session on a five-point scale from novice to expert. It never looked at job titles — it read behavior. Anthropic reports the classifier's agreement with independent telemetry exceeded 90%.

The occupational result is the one to anchor on, because it's the one most people get wrong in both directions. It is not true that "non-coders are now as good as coders." And it is not true that the gap is large. On verified success — which required both the classifier judging the session a success and independent telemetry confirming it through passing tests, commits, or explicit user confirmation — software engineers reached about 34% and non-software occupations about 29%. A real five-point gap, but a narrow one. Every one of the ten largest occupation groups, Anthropic reports, landed within seven points of software engineers.

Here is the interpretation worth holding onto. The technical-versus- non-technical binary has stopped working as a performance predictor. When the agent handles syntax, the remaining variable is judgment about what to build and whether the result is right — and that judgment is distributed across professions, not concentrated in one. The narrowness of the gap is exactly why a non-coder should feel invited in, and the existence of the gap is exactly why a non-coder shouldn't expect the agent to think for them.

Anthropic's classifier didn't guess at expertise from a résumé. It measured three concrete things in each transcript: how precisely the user framed the work, what the user asked Claude to verify, and the direction of correction — whether the user corrected Claude or Claude had to correct the user. Crucially, the measure is task-specific, not credential-specific. A senior engineer asking their first Rust question registers as a novice for that session; an accountant who specifies reconciliation rules precisely and catches the edge case registers as an expert, regardless of whether they know Python.

That makes the three signals the most useful artifact in the entire study, because you can turn them into a self-test. Before you send your next prompt, score yourself against all three. None of them require knowing a programming language. All of them require knowing your own domain — which, if you're reading this for your field, you already do.

The study describes the classifier's signals but never publishes them as a user-facing checklist. The table below is our own: it maps each of the three signals — plus the session-structure data on how experts steer through trouble — into a concrete contrast between novice and expert behavior, and names what each shift unlocks inside the agent. None of the left-hand column requires you to be a novice forever. Every row is a behavior you can change on your very next prompt.

The leverage these behaviors unlock is measurable. Anthropic reports that expert-rated users triggered around 12 Claude actions and ~3,200 words of output per prompt, while novices triggered about 5 actions and ~600 words. That is a 2.4x gap in actions and more than a 5x gap in output volume — from the same single instruction. The expert isn't typing faster. They're front-loading enough clarity that one prompt does the work of several.

The most actionable thing in the entire study is also the most easily missed: Anthropic states that most of the gain is concentrated at the novice-to-intermediate transition. The jump from novice to intermediate is larger than the additional jump from intermediate to expert. Verified success climbed from about 15% for novices to the 28–33% range for intermediate and expert sessions — meaning the bulk of the improvement happens the moment you stop prompting like a beginner.

That reframes the whole project. You do not need to become a world-class agentic operator to get most of the value. You need to cross one threshold: from vague to specific, from accepting output to checking it, from abandoning when stuck to redirecting. Becoming "intermediate" is not a years-long apprenticeship — it's a handful of habits drawn directly from the three signals. For the prompt-level patterns that move you across that line, our review of which prompt patterns actually worked in H1 2026 is a useful companion.

The study reports these numbers across several different sections; no single table in the paper or the coverage collects them in one place. Below we assemble the tier-by-tier picture so the "get to intermediate" argument is visually obvious. We publish the success and abandonment figures Anthropic states for each tier, and mark the action and output columns as expert/novice only — because Anthropic publishes those two metrics for the endpoints, not for the intermediate tier. Every number is from Anthropic's classifier and is vendor-stated.

Read the abandonment column alongside the success column. When sessions ran into trouble, Anthropic reports that expert users recovered to success about 15% of the time versus 4% for novices, and that novices abandoned struggling sessions at 19% against 5–7% for intermediate and expert users. The resilience gap matters as much as the headline success gap: a lot of "failure" is really just giving up early. Knowing how to steer a stuck agent — which is signal three, correction direction — is what keeps a hard session alive.

One result is too instructive to skip: management occupations ranked highest on verified success, slightly above software engineers. Read it the right way and it's a gift, because it tells you what mental model the work rewards. Translating a goal into precise requirements, judging whether an output meets the bar, managing iteration when the first attempt misses — that is the daily job of a good manager, and it maps almost exactly onto the three expert signals. If you've run projects or people, you already own the mental model. Your challenge is applying it to an agent, not learning a new skill from scratch.

One honest caveat, which Anthropic itself flags. The management edge may partly be a measurement artifact: managers are more likely to explicitly confirm success in the transcript, and the classifier rewards that explicit confirmation. So the ranking might reflect genuine skill transfer, a quirk of how the metric is read, or both. Don't treat "managers beat engineers" as a clean, settled finding — treat it as a strong hint about which habits travel well into agentic work.

Here is the practical translation — a routine that operationalizes the three signals plus the resilience finding. Run it on your next AI build, whether that's a report, an automation, a small tool, or a web page. Each step pairs a finding from the study with the concrete move it implies.

This is the practical core of how we run an AI digital transformation engagement: not handing a team a tool and hoping, but teaching senior practitioners to brief, verify, and steer agents inside their own domain. The same discipline underpins how we build production software — the person steering the agent is the leverage point, and the workflow is designed around their judgment. For broader context on the tool landscape and how teams adopt it, see the H1 2026 AI coding retrospective and the 30-person dev shop adoption case study.

A forward-looking read is worth stating directly. As of April 2026 the expertise gap had not narrowed — Anthropic flags this as a trend to watch, not a temporary artifact. If that holds, the advantage doesn't accrue to whoever buys the best model; it accrues to whoever brings the sharpest domain judgment to it. For a non-coder, that's genuinely good news: the moat is built from the expertise you already have, and the only thing standing between you and an intermediate-level session is a more disciplined brief.

A single vendor study, however large, is not the final word. Four caveats keep the conclusions sober. First, the success rates come from Anthropic's own proprietary classifier and have not been independently replicated — read them as "Anthropic's data shows," not as settled fact. Second, the study excludes Cursor, VS Code integrations, headless sessions, and Agent SDK usage, so the population skews toward users of Anthropic's own CLI, claude.ai, and desktop app; the patterns may not generalize one-to-one to Cursor or Windsurf users.

Third, the management-beats-engineers result carries Anthropic's own measurement-artifact caveat and should be read as a directional hint, not a clean finding. Fourth, the study's figure for the average session's economic value rising about 27% is a relative benchmark estimated against public freelance posting rates — the researchers call the direction reliable but the specific percentage measurement-noisy, so we treat it as a trend toward higher-complexity work rather than a dollar figure. None of this undercuts the core finding. It simply means the right posture is to treat the study as strong directional evidence, run your own measurements on your own work, and let the pattern — not any single percentage — drive how you prompt.