Half of them had never made anything with AI. Four had never written a line of code, not even with Scratch. Five days later, every team had a working product on a public URL, demoed live to company leadership.
Abelium Summer Jam 2026 — AI Group ran full working week in July 2026, 08:30 to 15:00, at Abelium offices in Ljubljana. Eleven participants aged 15 to 19 formed three teams, supported by six rotating mentors from Abelium and partner companies.
Why we ran it
This was Abelium's own initiative, and it had a thesis to test: that any young person — across a wide range of ages and starting points — can be brought up to speed through hands-on experiential learning, to the point of using AI effectively to build real digital tools.
That belief is not abstract for us. Abelium is an active member of the AmCham Slovenia committee »Pripravljeni na prihodnost« (Ready for the Future), represented by dr. Boris Horvat, Director of Abelium. The committee has argued for years that computer science belongs in Slovenian schools as a compulsory subject, taught vertically from the first grade — it said so in 2022, and it said so again in May 2025, responding to the amendments to the Primary School Act that introduce the subject Informatika in digitalne tehnologije.
The committee welcomed that step and named its limits plainly: the subject arrives in 2028, it begins in grade 7, and as drafted it points at using digital tools rather than understanding how they work — programming, data, critical judgment. The numbers behind the concern are not comfortable reading. Under ICILS 2023, 11% of Slovenian primary pupils use ICT daily for schoolwork, the lowest share among all participating countries. 0.6% of teaching time in primary and secondary school currently goes to computing. The Slovenian economy needs at least 5,000 ICT professionals a year.
Abelium Summer Jam is the other half of that argument. Making the case in a committee is one thing; showing what becomes possible when young people are handed professional tools, a real deployment pipeline and mentors who unblock rather than lecture is another, and it is harder to argue with. Fourteen months after that response was published, eleven teenagers spent a week doing exactly what the committee says schools should be preparing them for — not using finished software, but building it. The second reason was simpler. We wanted teenagers — and their families — to see what engineering at Abelium actually looks like from the inside.
The design: an environment, not a curriculum
The teaching instrument was not a lecture series. It was a designed environment: professional AI tools, a real deployment pipeline — GitHub, Vercel, TypeScript, Next.js — defined team roles, a transparent scoring system, and mentors briefed to unblock rather than instruct.
The week had an arc. Monday was an individual play day built around a 42-card challenge deck, where every card ended with the same real question: which Claude tool fits this best? Tuesday teams formed, picked a problem worth solving, and wrote the plan into the shared repository's ReadMe. Wednesday the roles kicked in and the building got serious. Thursday was for going live and showing the work to real people. Friday they demoed to company leadership.
Milestones were deliberately day-agnostic. A team that deployed on Tuesday earned exactly the same points as one that deployed on Friday morning — the calendar was a rhythm, not a gate.
Each team ran four defined roles: the project manager, who owned the plan and scope; a backend and a frontend coder, who split the build; and a designer, who owned look, feel and the pitch. Designer and frontend coder worked on wireframes at the same time rather than in sequence, which removed the classic hand-off bottleneck. Only one person drove Claude Code at a time, and the seat rotated — so everyone learned to build, and nobody spent the week in version-control chaos.
Three Claude surfaces, three muscles
Using all three Claude surfaces deliberately is what separated the week from a generic build-with-AI camp.
- Claude chat trained the strategist — research, problem validation, business thinking.
- Claude Design trained the creative — branding, UI, pitch visuals.
- Claude Code trained the builder — the actual working product.
Every participant worked inside Abelium's Claude Team plan on a Standard seat, with their own account rather than a shared demo login. Two seats were upgraded to Premium partway through the week, once it was clear that the heaviest build work needed the extra tokens.
The durable takeaway was never the tool. It was the loop — plan → prompt → review → judge → iterate — and the judgment to know when to trust the output and when to push back on it. Mentors modelled that loop constantly instead of lecturing it. The rule of the week was simple: stuck for fifteen minutes, ask.
The platform underneath, built with Claude Code
The scoring system that drove the week runs on software Abelium built for it — itself a demonstration of the week's message.
A Firebase backend on Cloud Functions and Firestore in europe-west3 treated every point as an immutable ledger event. It was built and maintained with Claude Code, by one engineer, alongside the rest of the preparation. Running it for the whole of July 2026 cost €0.09 in cloud charges.
Points paid for courage and shipping rather than talent. The Half Rule guaranteed that an honest attempt earned half points even when it failed, with a mentor's one-sentence reason attached as human feedback. The TV read only published totals, which was announced at the 14:30 ceremony.
What shipped
[KibbleBee], by 4 Tokens — a personalised dog-feeding planner. Enter your dog's profile, or start from Rex the golden retriever, and get a calculated feeding plan.
[SplitFlik], by Push*pull*legs — bill-splitting without napkin math. Scan a receipt, tap who ordered what, and the app computes who owes whom, with groups, QR invites and one-tap settlement. Presented with a freemium model: free with ads, premium at €3 a month.
[PH-Slo], by Miza — the nearest working hydrant finder, a tool for fire brigades. One tap marks where the fire is and shows the nearest hydrant, with road-based navigation and the data that matters at a glance: above or below ground, diameter, flow rate.
All three were demoed live from public URLs at the Friday showcase, opened by the audience over QR codes.
Results
Three teams entered. Three teams shipped. Final standings: 4 Tokens 175 points, Push*pull*legs 160, Miza 143 — a close race, which is what the scoring system was designed to produce.
Where they started. In a poll taken before the week began, half the group said they had never created anything with AI — they had only chatted with it, or had never used it at all. Four participants had never programmed at all, not even in Scratch. Some had coded before Summer Jam; none had taken a product of their own to a public URL.
Where they finished. Every team had a deployed, publicly reachable product and had presented it live. Ten of the eleven were there on the final day — one had fallen ill — and all ten answered the closing survey. Asked how satisfied they were with what they had built, seven of the ten answered 4 or 5 out of 5.
The magical wand of Claude. Day-one challenge cards were the ice-breaker for those who never coded — working mini-apps built and shown to a mentor within the first day: Tic-tac-toe, flashcards, a memory game, hangman, 2048, Snake, Connect Four, a logo, a bill splitter. They created it in various Claude tools and got the needed self-confidence they can do it.
Asked what they would remember most, participants wrote:
- "Claude."
- "How useful Claude is."
- "That Claude can do anything, if you tell it the right thing."
- "That we worked together really well."
— anonymous end-of-week survey, translated from Slovenian.
What happened afterwards. In the weeks that followed, several participants told their mentors they had carried on building things of their own with their Claude accounts. We did not measure that, and we are not going to dress it up as a metric — it is simply what they said, and it is the outcome we were actually after.
What it taught us
On Friday afternoon the mentor team held a structured retrospective, and it was not uniformly flattering. We are publishing what it found, because the useful part of a pilot is the part that didn't work.
Students were overwhelmed by the number of new technologies at once. A disproportionate share of mentor time went to debugging project plumbing — framework, hosting and database wiring — rather than to substantive mentoring. One team fell significantly behind on deployment issues alone. The end-of-week survey agrees from the other side: asked what was hardest, participants named the frontend and getting a first idea; asked what mentors helped with, they named the frontend, Vercel, and "figuring out what doesn't work."
Monday was socially quiet. Opening up was gradual. This generation communicates less readily face to face, and ice-breakers, a physical activity and digital facilitation tools belong in the program as first-class elements, not afterthoughts.
Ideation stalled on self-censorship and premature judging — including prematurely outsourcing judgment to AI, which is precisely the habit the week was meant to break. Participants asked for exactly this in the survey: more brainstorming before settling on one idea.
The honest conclusion: a scoring system is scaffolding for motivation and feedback, not a substitute for pedagogy. It does its work only when embedded in deliberate onboarding, facilitated ideation, social warm-up and trained mentorship.
What to do with this pilot case
A pre-configured project template with a working deployment, so every team starts at feature level instead of at plumbing level. A moderated brainstorming hour per team. Ice-breakers and a physical activity early in the week. Pre-assigned roles. Mentor training and shared mentoring standards.
Every component of the program — the weekly arc, the challenge deck, the role system, the scoring rubric, the points application, and the full set of printed and projected materials — exists as a reusable artifact. That is what makes it repeatable, and this makes this pilot reusable anywhere.
The formula we will keep testing: real tools + real stakes + designed motivation + mentors who unblock = teenagers who can build with AI.
One week run by one company is not an education policy, and we are not pretending otherwise. But it is evidence, gathered in practice rather than argued in principle, and we will keep putting it on the table — in the Ready for the Future committee and anywhere else it is useful. If a group that included people who had never written a line of code could ship working software in five days, the question of what Slovenian schools could do with thirteen years of schooling is worth asking harder.
