What I Learned from Anthropic’s Hackathon Winners
Anthropic ran a hackathon in February. Thirteen thousand people applied; five hundred got in; one week to build something with Claude Code; five winners.
Four of the five didn’t have CS degrees. The headlines wrote themselves.
“Coding is dead.” “The lawyer who beat 500 developers.” “You don’t need to know programming to create solutions” — that last one is a direct quote from the third-place winner, Dr. Michał Nedoszytko, a Brussels cardiologist whose PostVisit app helps patients understand their visits.
The framing is irresistible: AI tokens + a good idea + a domain expert = winner. Coding is over. Anyone can build now.
Schooly Mammoth is built on the idea that real work experience is what makes a builder — and you can’t get enough of it in a classroom. The hackathon headlines read like a counterexample: AI plus a domain expert was apparently enough, no work experience required. So I went and looked at the winners. What I found supports the work-experience premise more clearly than the headlines do.
The doctor has been programming for twenty years.
Dr. Nedoszytko is, by his own GitHub bio, “cardiologist/software developer.” He’s been building healthcare software alongside his clinical practice for two decades. He founded previsit.ai in 2023 — meaning the hackathon submission wasn’t a from-scratch miracle; it was the third year of an existing product line. He pushed 349 commits in seven days while coding on a flight from Brussels to San Francisco. You don’t push 349 commits in seven days unless you’ve done it before, a lot.
His “you don’t have to know programming” line is the perfect quote for the headlines. It’s also misleading — not in what he meant, but in what gets done with it. He clearly knows how to program. He just didn’t get there through a CS degree.
The same is true of every other winner.
Same pattern, all five.
Mike Brown — first place, the personal injury lawyer who built CrossBeam to handle California ADU permit corrections. His public GitHub includes a project called agent-skills-public-mike, described as “open-source AI agent skills for small businesses… part of PBC OS.” He’s been shipping reusable AI agent infrastructure for some time. His GitHub username mikeOnBreeze hints at a previous project named OnBreeze — likely the AI client-intake system he built into his own law practice a year before the hackathon. CrossBeam wasn’t his first rodeo.
Jon McBee — second place, the engineer who built Elisa, a visual programming environment, for his 12-year-old daughter. He has dozens of public AI-tooling repos: multi-agent orchestration CLIs, agent collections, even a San Diego Permit FAQ Bot — meaning he was already thinking in CrossBeam’s exact problem space before the hackathon started. He produced 39,000 lines of code with 1,500+ tests in 30 hours. That throughput isn’t a beginner’s miracle. It’s compounding velocity from years of prior practice.
Kyeyune Kazibwe — fourth place, the road infrastructure specialist in Uganda whose dashcam-to-investment-appraisal tool, TARA, won the “Keep Thinking” prize. His GitHub bio reads, literally, “Building AI tools for Civil Engineers.” He maintains an open-source library called OpenEng — sixty-five engineering skills across thirteen domains. He has a portfolio of agent-memory tools, autonomous research systems, and meta-prompting frameworks. He didn’t show up cold to a hackathon.
Asep Bagja Priandana — fifth place, the electronic musician whose Conductr turns Claude into a live virtual bandmate. Fifteen years of self-taught engineering. Founder of Nanas Sound, a music software/hardware company in Estonia. Earlier products include AirBending, a gesture-based MIDI controller marketed as “your hands are the instrument.” He’s been on this for over a decade.
Five winners. Five long building practices.
So what actually changed?
Two things.
One: what counts as programming is shifting. Programming used to mean typing syntax — declarations, loops, edge cases, framework choices. AI is taking over the typing-the-syntax part. What’s left is the harder part: directing the work, architecting the system, knowing when the model is wrong, owning the output, judging when it’s done. That’s still programming. It’s a higher-leverage version of programming — and higher-leverage tools have been arriving for programmers for decades, from assembly to C to Python to web frameworks to IDEs. AI coding is the biggest leap yet, but it’s the same direction. A lawyer with twenty years of legal-domain expertise plus this new kind of programming practice is more dangerous than either alone.
Two: the entry cost dropped. Building practice used to require a CS degree, a laptop you could afford to brick, a shelf of books, and a mentor or a job that let you write production code. AI didn’t remove the practice — it removed several of the gates that used to block it.
When the headlines say “you don’t need to know programming,” what they actually mean is “you don’t need a CS degree to start.” Those aren’t the same claim.
A 16-year-old with a real problem, an AI subscription, and a mentor can now log build-hours that would have been functionally impossible ten years ago.
Practice, not magic or major.
The lesson the headlines are missing has two parts.
Not magic — the tool isn’t a cheat code. AI tokens plus an idea, with no practice, doesn’t win a hackathon. The five winners aren’t evidence that practice is unnecessary; they’re evidence that practice can compound much faster now than it used to. McBee’s 39,000 lines in 30 hours is impressive. It is not miraculous. It is what twenty years of building, with the latest tools, looks like.
Not major — whatever you majored in (or, if you’re 16, whatever you’re considering majoring in) doesn’t determine whether you can build. Lawyer, doctor, civil engineer, musician, biotech engineer — five different starting points. What links them isn’t a major. It’s that all five, regardless of the path they took, ended up with both a domain they understand at the marrow level and enough programming practice to ship.
That combination is the whole game now: deep domain knowledge plus practiced AI-augmented programming, applied to a problem you actually have.
When do you start practicing?
I’m a coach. I’ve helped athletes and nerds win. None of them got there without reps. Thousands of them. The day of the competition doesn’t make you; the years before do.
AI didn’t change that. AI changed who can put in the reps and how fast they can compound.
The hackathon winners had to scrape their reps together at night, on weekends, on transcontinental flights, after a degree, often in their thirties and forties. They got there anyway. But imagine the version where you don’t have to wait that long.
That’s the idea behind Schooly Mammoth. We give 16-year-olds the practice these winners had to assemble themselves. Real products. Real users. Real teammates across pathways. School credit. Mentor feedback. By the time a Schooly Mammoth intern graduates high school, they’re not first-timers. They’re four-product-deep builders. By 20, they’ll have logged the build-hours these winners didn’t get to until their thirties.
The first-place winner, Mike Brown, said the line that should be the title of every CTE program in the country:
“You need to build to learn. I wouldn’t be here with CrossBeam if I hadn’t done all that other stuff first.”
The “other stuff” is the practice. AI doesn’t replace it. It just lets you start sooner.
That’s why we built Schooly Mammoth.
Leave a Reply
You must be logged in to post a comment.