Why This Skill Matters Now

Two things are true at once, and most of the internet only ever tells you one of them. Shipping something that works got dramatically cheaper. Shipping something good did not. Every number below carries its source, because you will be arguing about this for the rest of your career.

Guess before you read

Make a prediction first. Not because guessing is fun, but because a number you committed to sticks, and a number you scrolled past does not. Three questions, one slider each. Commit to a number before you press reveal.

The 95% number is real, and it is not what you think

In March 2025 one sentence did more for vibe coding’s reputation than every demo combined. Jared Friedman, a managing partner at Y Combinator, said that for about a quarter of the startups in YC’s Winter 2025 batch, roughly 95% of the codebase was AI-generated (TechCrunch, 6 March 2025).

Read it again slowly. It is 25% of one batch, not 95% of Y Combinator. The 95% compares human-typed to AI-typed characters, and it excludes imported library code. And Friedman’s own framing is the part that never travelled: those founders were highly technical people who could have written every line by hand. In the same YC video Diana Hu said judging model output takes taste and training, and Garry Tan noted that the reasoning models of the day were poor at debugging.

So the stat is true, and what it means is this: experienced builders, given a good tool, stopped typing. It does not mean the typing was the hard part.

The best trial we have said the opposite

METR is a research group that runs evaluations on AI systems. In early 2025 it ran the trial everyone now argues about, published on 10 July 2025.

Sixteen experienced open-source developers. 246 real issues from their own repositories — mature projects averaging over a million lines of code, which these developers had worked on for around five years. Every issue was randomly assigned to AI-allowed or AI-disallowed. The tooling was mostly Cursor Pro with Claude 3.5 and 3.7 Sonnet.

With AI, the tasks took 19% longer.

The part that matters more than the result: before starting, those developers forecast they would be 24% faster. After finishing — after actually living through every task — they still believed they had been about 20% faster. The stopwatch said 19% slower. That is roughly a 40-point gap between the feeling and the measurement, and the people in the gap were experts on their own code.

What happened to that result afterwards

Evidence that never updates is not evidence. On 24 February 2026 METR published the rerun: 57 developers (10 returning from the original study plus 47 new), more than 800 tasks across 143 repositories, starting in August 2025.

Returning developers: 18% less time. Confidence interval −38% to +9%. New recruits: 4% less time, interval −15% to +9%. Both intervals cross zero, so neither is a number you can bank.

Then METR did something rarer than running a study: it called its own dataset an unreliable signal and started redesigning the method. Some developers refused to take part at all if it meant working without AI. Between 30% and 50% of participants quietly avoided submitting the tasks they least wanted to do by hand. METR’s own reading is that its estimate is likely a lower bound — the best AI use cases selected themselves out of the experiment.

The honest sentence today: the sign has probably flipped; the size is unknown.

In May 2026 the same lab asked 349 technical workers what they felt. Median self-report: 1.4–2x more value, 3x more speed (METR, 11 May 2026). METR attached its own warning to its own survey, pointing back at the trial where people overestimated by about 40 points.

Everyone uses it. Almost nobody trusts it.

Stack Overflow’s 2025 developer survey collected more than 49,000 responses across 177 countries and was released on 29 July 2025. 84% of developers use AI tools or plan to. In the same survey, 46% distrust the accuracy of what those tools produce, and only 3.1% trust it highly. The top frustration, named by 66%, is an answer that is almost right but not quite; 45% say debugging AI-written code takes longer (survey.stackoverflow.co).

That is not a contradiction. That is a profession that has decided the tool is worth using and worth checking.

One more thing about that survey, and it is a skill rather than a fact: there is no published 2026 edition yet. Plenty of articles right now run the 84% and the 3% under a “Stack Overflow 2026” headline. Those are the 2025 numbers wearing a new date. Check the year on every statistic anyone hands you, including the ones on this page.

Code quality points the same way. GitClear analysed 623 million code changes from 2023 to 2026: copy-pasted lines rose from 9.4% to 15.7% of new code in the first half of 2026, duplicated blocks rose about 81%, and refactoring — tidying what exists instead of adding more — fell from roughly 21% of changes in 2022 to 3.8% (GitClear). GitClear sells code-quality tooling and the analysis is correlational, so take the direction, not the decimal. The direction is: agents add. They rarely tidy, unless you ask.

Security has not improved on its own either. Veracode’s July 2026 report found 56% of generated samples passing a security test, so 44% introduced a known kind of flaw — against a 55% pass rate a year earlier (Veracode). Its March 2026 edition put syntax correctness above 95% (Veracode). Models learned to write code that runs. They did not learn to write code that is safe. Those test prompts contain no security instructions, so the number measures the default, not the ceiling. That is the good news: the default is the part you can change.

The money says this is not a fad

• On 14 August 2026 SpaceX filed a Form 8-K recording that Anysphere, the company behind the Cursor editor, had merged into a SpaceX subsidiary, all-stock, at an implied equity value of $60.0 billion. A code editor, absorbed by a rocket company.
• Lovable said in June 2026 it had reached $500 million in annualised revenue, with a million new projects started every week and over 50 million built in total (TechCrunch, 9 June 2026) — reported alongside the question nobody outside the company can answer: how many of those are still alive?
• Anthropic said in February 2026 that Claude Code was running above a $2.5 billion annual run-rate, more than double where it began that year (Anthropic).
• JetBrains surveyed over 15,000 professional developers between May and July 2026: 90% use an AI coding agent at work at least weekly, 68% daily. Claude Code reached 39% work adoption, up from 18% in January 2026 (JetBrains Research).

Argue about whether any one valuation is sane. You cannot argue that this is a niche — nine in ten professional developers use an agent at work every single week.

The floor rose. The ceiling did not move by itself.

Two of those numbers say the work got easier. Two say it did not. Put them side by side and one sentence falls out.

The floor rose enormously. Things that used to need a team, a budget and half a year now need a weekend and a clear head. A million new Lovable projects a week is the floor rising, and it is why you are reading this instead of a textbook.

The ceiling did not move by itself. Faros AI looked at telemetry from more than 10,000 developers on 1,255 teams: the teams using AI heavily merged 98% more pull requests — and spent 91% more time reviewing them. Pull requests grew 154% larger, bugs per developer rose 9%, and there was no measurable improvement at the company level (Faros AI). The work did not vanish. It moved, from writing to reading.

One builder counted it on his own project. In March 2026 Mac Budkowski described getting a working prototype of a small Farcaster app in about an hour, then spending more than 100 hours making it shippable: interface, infrastructure, contract security, edge cases (kanfa.macbudkowski.com). The first hour is the floor. The hundred hours are the ceiling.

The Science study of 160,000 GitHub developers found the same shape from the other end: less experienced developers used AI on more of their work (37% versus 27%), and the measurable output gains went almost entirely to the experienced ones.

Anthropic ran the version that should concern a student most. In January 2026 it put 52 mostly-junior engineers in front of an unfamiliar Python library. Those who delegated to AI scored 50% on a comprehension quiz afterwards; those who wrote it by hand scored 67%. The AI group finished about two minutes faster — not a meaningful difference. The people who kept their scores asked the model to explain while it worked, instead of only to produce (Anthropic Research).

The floor is a gift; take it. The ceiling is a skill, and nobody hands it to you. The gap between feeling fast and being fast is the single most useful thing in this course, and every module after this one is a technique for closing it.

Closing the gap starts with one prompt

The perception gap has a practical cause: when an agent finishes you see a confident summary, not the list of things it never checked. So ask for that list. Paste this at the end of any agent session, before you accept anything.

Before I accept this work, answer these four questions. Do not change any code yet.

1. List every file you changed, one line each, saying what changed.
2. For each change, say how you verified it actually works.
   If you did not run it or test it, write "not verified".
3. List what you assumed about my code, my data or my environment
   but did not check.
4. Name the single change most likely to be wrong, and say why.

Run it once today on anything you have already built with an agent. Question 2 is where the surprises live, and one honest “not verified” is worth more to you than this entire page of statistics.