Anthropic Just Studied 400,000 Claude Code Sessions. The ‘Learn to Code’ Era Is Over.

Picture a 44-year-old immigration lawyer. She spends 524 hours with Claude in 2025. She has never written a line of code. By February 2026 she is running a solo law practice on software she built herself, plus a SaaS platform for legal workflow automation, and she carries a Chief Legal Engineer title at an AI startup.

A year ago, that story was a fantasy. Anthropic just dropped the paper that explains why it isn’t anymore.

400,000 sessions. 235,000 people. Seven months of data.

A software engineer ships verified work with Claude about 30% of the time. A lawyer doing the same coding task ships about 26%. The gap has not widened in seven months. It is also not closing.

The gap that ends an era is five points wide.

I will walk you through the 9 findings, the prompt for each, and our lawyer’s path as the proof.

Part I: The 5-Point Gap That Ends an Era

Anthropic measured every coding session by occupation. Software engineers and other computer-and-math workers reached verified success in about 30% of their sessions. Users from every other profession reached verified success about 26%. Inside seven months of data, that gap has not widened. It also has not closed. Anthropic notes that every major occupation in the dataset lands within seven points of software engineers.

Our lawyer sits in the 26%. She is also what the top of that distribution looks like in real life. Verifiable, working code. Real product. Paying customers. None of it written by a software engineer.

The “learn to code” pitch has been sold for fifteen years on a single premise: you are locked out of this work until you study CS. Anthropic just measured what that premise actually costs you. A few points.

The bootcamp industry has been quietly contracting in the background. BloomTech, the school that used to be Lambda, cut roughly half its staff. Codeup closed. Launch Academy paused enrollment. Dev job postings on Indeed have sat well below pre-pandemic levels for two years. The industry that sold protection inside this gap has been thinning out for a while. Now there is a number on the gap.

The asset for this part is the simplest one in the whole article. Open every Claude Code session with this in the first prompt:

Four lines. That preamble, just translated into immigration law, is what our lawyer opened every session with. Enough to close most of the five-point gap before Claude writes a single import statement.

Part II: Domain Expertise Doubles Verified Success

Inside that flat lawyer-vs-engineer number, there is a steeper one.

A session where the user looks like a novice at the task reaches verified success 15% of the time. Intermediate reaches 28%. Expert reaches 33%.

The jump from novice to intermediate is bigger than the jump from intermediate to expert.

A working grasp of your own domain captures most of the gain. Deep specialization adds a little on top. You do not need to be a 20-year senior. You need to actually understand the problem you are asking the agent to solve.

Our lawyer spent twenty years in immigration law before she ever opened Claude. Her 524 hours were not her learning to code. They were her teaching Claude immigration law. Every session was an expert session in her own field, by the very definition Anthropic uses.

This is what makes the bootcamp business model collapse. Bootcamps spent fifteen years teaching you the syntax. Anthropic’s data says the syntax is the cheap part. The expensive part is knowing what you actually want.

The asset for closing the novice gap is one prompt. Drop it on any task before you start asking Claude to do the work:

Novices skip this. Experts run it by reflex. One prompt is the entire gap.

Part III: The 70/80 Division of Labor

Anthropic built a classifier that splits every session into planning decisions (what to do, what counts as done) and execution decisions (which files to change, what code to write, what to run).

The user makes 70% of the planning decisions. Claude makes 80% of the execution decisions.

The split is sharp. Stable across occupations. Does not blur as the models get better.

People decide what to build. The agent decides how to build it.

Andrej Karpathy walked onto the YC AI Startup School stage on June 17, 2025 and called this new layer “Software 3.0.” Prompts are the new source code. English is the new programming language. LLMs are the new CPUs. He predicted the architecture out loud almost a year before Anthropic measured it.

Most of our lawyer’s 524 hours were planning. She would describe an I-130 family-based petition workflow in immigration-attorney English, and Claude would execute the implementation in Python and TypeScript. She kept the planning. She delegated the execution. The 70/80 was her entire architecture.

If you treat the agent like a junior coder who needs babysitting, you keep the execution decisions, the agent runs at half power, and your session ends in the abandoned 5-7%.

The asset for forcing the split is a 3-line spec, dropped at the very start of any session:

This is the manager move. Anthropic’s data says it ships 30-33% of the time.

Part IV: Experts Pull 5x the Work per Prompt

Every prompt the user sends sets off a chain of actions Claude takes on its own.

A novice prompt averages 5 actions and 600 words of output.

An expert prompt averages 12 actions and 3,200 words of output.

Same tool. Same model. Same minute. Five times the work, off the same number of keystrokes.

The variable is what loads onto the prompt: the domain context, the constraints, the prior decisions, the test that defines done. When the load is rich, Claude has a 100-step problem to chew on. When the load is empty, Claude has a fill-in-the-blank.

The prompt is the proxy for how much you actually know.

Our lawyer’s 524 hours were the slow accumulation of that load. By month six, every prompt she sent had years of immigration-law context behind it: the USCIS form numbers, the priority dates, the consular processing quirks, the affidavit-of-support thresholds. Every prompt ran 12-action chains because every prompt was loaded.

The asset for unlocking the 12-action chain is a context-loading preamble. Paste it once at the top of any project, then refer to it every time:

This is the spec that makes the next prompt 12 actions instead of 5.

Part V: Debugging Is Dying. Operating and Writing Are Rising.

What people do with Claude Code changed hard in those seven months.

The share of sessions spent fixing broken code fell from 33% to 19%. Almost cut in half.

In its place, operating software grew from 14% to 21%. Writing documents and analyzing data roughly doubled, from about 10% to 20% combined.

If you are still using Claude Code to debug, you are spending your time on the slice that just halved.

The growth is in everything that is not writing functions. Running pipelines. Monitoring deploys. Pulling a dataset and writing the memo on top of it. Drafting a 40-page report. The agent does fewer and fewer “fix this broken function” jobs and more and more “operate this entire workflow” jobs.

Our lawyer barely touched debugging. She never had the muscle memory to lose in the first place. Her sessions were build-from-zero from day one, then operate-and-extend after that. She moved with the curve, because she had no other curve to follow.

The asset for this part is just a list. Five non-coding workflows the data says Claude Code now handles cleanly:

Each takes about one prompt and one verification turn. None of them is “coding.” All of them ship.

Part VI: Managers Verify Higher Than Engineers

There is one occupation in the dataset that beats software engineers on verified success.

Management.

Not by a lot. But in coding sessions, managers land slightly above software and math on verified success rate. Anthropic flags the obvious caveat: managers may simply confirm in chat when something is done, which boosts the “verified” signal. Fair. But the operational read is harder to dismiss. The skill that runs an agent successfully is the skill that runs a person successfully.

Set the goal. Define done. Accept the path. Check the work.

Our lawyer did not become a software engineer. She became a Chief Legal Engineer, which is the manager title for a person who runs a team of agents instead of a team of people. The job exists because the data does.

A manager treats Claude Code like a direct report. An engineer treats it like a junior who needs supervision. The first one finishes the sprint.

The asset is a manager-style frame to open any non-trivial session:

This is the prompt that converts a session into a sprint.

Part VII: The 4x Abandonment Cliff

Among sessions that hit trouble, novices abandon 19% of the time. Everyone else abandons 5-7%.

Novices do not fail more. They walk away more.

Expert sessions get into the same fires. They climb out. The difference Anthropic could measure was not the fire, it was the recovery.

That is what 524 hours buys. The first hundred hours our lawyer abandoned constantly. By hour two hundred she was climbing out of every fire. The recovery muscle is the same muscle you build going from intermediate to expert: not more skill, more stubbornness wrapped around domain context.

The cheapest gap in the whole study to close is one prompt. When a session derails, stop free-prompting. Open this:

This is the “pull up out of a flat spin” prompt. Drop it instead of telling Claude to “try again.”

Part VIII: Task Value Rose 27% in 7 Months

Anthropic approximated the dollar value of every session by mapping it to comparable freelance job postings, calibrated against a public dataset. The estimates are coarse, but the direction is the point.

Average session value rose 27% from October 2025 to April 2026.

The climb was not uniform. Building tasks grew about 43%. Operating software grew about 34%. Fixing tasks grew about 32%. Every kind of work got more valuable per session, in seven months.

The value is leaving fixing and entering building.

The agents got better at the work upstream of code. The hourly rate the work would clear on a freelance market moved with it.

Our lawyer spent month one fixing. Month four operating. Month seven through twelve building. She moved through the curve in the same order the entire market did, and she ended up where the value rose hardest. The Chief Legal Engineer title is what +43% looks like on a single person.

The asset is a directional bet. Three task categories where Anthropic’s data says the value rose hardest:

Treat it like a portfolio. Spend more of your hours on (1) and (2). Take (3) only when the unit price is high.

Part IX: The 4-Turn Session

The whole pattern lives inside a four-turn session.

A typical Claude Code session is four user prompts. Each prompt sets off about 10 actions on average, sometimes over 100. Claude writes about 2,400 words of output per turn.

Four prompts. Forty actions. Ten thousand words of output. That is the unit.

Go in without a plan, and you will spend all four turns on negotiation and end with nothing committed. Go in with a four-turn structure, and you ship. Our lawyer’s 524 hours converged on something close to this exact shape by month four. She stopped having “exploratory chats” with Claude and started running these four turns, hundreds of times.

Here is the template the data implies:

The entire system is four prompts wide.

The Take

Andrej Karpathy named “vibe coding” on February 2, 2025. By the end of the year, Collins Dictionary had made it the Word of the Year. In March 2025, Dario Amodei said AI would be writing 90% of code within three to six months. The forecast moved closer to reality every quarter. The bootcamps spent 2025 quietly renaming themselves and pretending they were “AI academies” now.

The Anthropic study is the receipt for what all of this does to the skill stack.

Coding agents made a coding background less relevant to programming. They made domain expertise the new bottleneck.

A five-point gap is not the end of programmers. It is the end of the credential. The senior engineer still wins on the margin. Everyone with a working grasp of a domain joins the field they thought they were locked out of.

The bootcamps were selling protection inside a gap five points wide.

The 22-year-old being told to “learn to code first” today is racing a 44-year-old immigration lawyer with 524 hours of Claude Code conversation, a Chief Legal Engineer title, and twenty years of immigration law sitting inside every prompt he sends. ** The lawyer is ahead.**

Everyone wants to win in the agent era. Nobody wants to learn anything outside of code.

Bookmark this & send it to whoever has been telling a 22-year-old to “learn to code first” since the day GPT-4 shipped.

More on AI and what actually wins in the agent era: @phosphenq