TechLabs Studio: Adventures in AI

Children shape personal AI Buddies with a collaborative group platform to build, debug, and share playable games and applications, learning to direct AI while preserving their creative spark .

Comment

Inspiration

I did not want to introduce children to AI through a lecture about what the technology is. I wanted them to experience what it means to create with it, and to notice who they become when a powerful tool is available to them.

There is an easy way for AI to weaken agency: the learner gives away the original idea, accepts the first output, and becomes a spectator. I wanted the opposite. Children should bring the impulse, make decisions, test the result, encounter bugs, change direction, and decide what is worth sharing.

That became Adventures in AI, a TechLabs course for learners aged 10 to 16, and TechLabs Studio, the platform that holds the course experience together.

What it does

TechLabs Studio is a bounded creation and sharing environment for young learners working with personal AI Buddies.

Each child enters through a fictional course identity rather than a personal email account. Learners shape how their AI Buddy should support them, then use conversation to turn an idea into a working browser-based creation. They test it, improve it, publish it intentionally, and share it with the course community. Other learners can open the creation, play it, and respond inside the course environment.

The platform connects several parts that usually live separately:

  • fictional learner identities and code-based access;
  • personal AI Buddy workflows;
  • a course Studio with bounded chat and sharing;
  • separate hosting for learner-created games;
  • a catalogue where the group can find and play published work;
  • a separate multiplayer path for creations that need it.

The separation matters. A learner device or AI Buddy should not need broad infrastructure credentials just to publish a game. Learner-created code should not run inside the Studio itself. Sharing should happen because the learner chose to share, not because every draft is public by default.

How we built it

The platform runs on Cloudflare. The Studio, game-serving origin, and multiplayer service are separated so each part has a narrower role and authority. The implementation uses TypeScript and Bun, Cloudflare Workers, D1 and R2-backed services, Durable Objects and WebSockets where required, Python components for bounded publishing workflows, and Playwright/browser tests for the learner journey.

The architecture was shaped around the real classroom path:

  1. A learner enters with a fictional Member Code.
  2. The learner connects a personal AI Buddy to TechLabs Studio through a scoped flow.
  3. The learner develops and tests a creation with the Buddy.
  4. Publishing produces a playable link without exposing infrastructure credentials.
  5. The learner shares the creation with the course group.
  6. Other learners open, play, and discuss it.

GPT-5.6, running through Hermes Agent, was the primary model I worked with for architecture, implementation, debugging, and test orchestration. Hermes provided the agent workflow, project context, and tool execution needed to work across the platform rather than treating each code change as an isolated prompt.

Codex in the ChatGPT app supported final validation and the deployment workflow across the classroom devices. I made the product, educational, safety, architecture, and deployment decisions, reviewed the generated work, and chose what entered the live course. TechLabs owns the project; I am its sole creator and developer.

Challenges we ran into

The hardest part was not generating a game. It was building a coherent system around AI-assisted creation for a real group of children.

Preserving the child’s creative role

An AI coding tool can produce something impressive very quickly. That speed can also erase the learner’s authorship if the system races ahead of the child’s idea. The Buddy and the course therefore frame AI as a collaborator that responds to the learner’s direction. The child remains responsible for the concept, the choices, the tests, and the decision to publish.

Giving useful capability without broad authority

The Buddy needs enough capability to help produce and publish a working creation, but it should not hold general deployment credentials. We separated identity, Studio access, game serving, and multiplayer responsibilities, then used scoped flows between them.

Running learner-created software safely

Children make unpredictable things. That is the point. It also means their generated games should not execute inside the same origin and authority boundary as the learner Studio. Separating game execution from the Studio reduced that risk and made the publishing path easier to reason about.

Making the classroom path work under real conditions

A system can pass a local test and still fail in a room full of learners using different devices. Student login, Buddy connection, long model responses, publishing, game links, and multiplayer all had to survive the actual course environment. Under deadline pressure, I repeatedly chose the smallest complete learner journey over unfinished administrative breadth.

Accomplishments that we're proud of

TechLabs Studio was used in a real Adventures in AI course, not only demonstrated as a concept. Learners created personal AI Buddies, built and debugged games, published playable links, explored one another’s work, and presented what they had made.

The result we care about is visible in their behavior. They were not watching an AI demonstration. They were making decisions, explaining their ideas, noticing limitations, and developing judgment about how to work with the technology.

What we learned

The children gave us the clearest evidence of what the course teaches.

One learner explained that a shorter, focused prompt was easier for the AI to understand, while a long and complicated prompt could introduce bugs and glitches. That is more important than producing a game in one minute. The learner was observing the system, forming a hypothesis, adjusting the instruction, and judging the result.

We also saw that bugs did not end the creative process. A moving character that could not be stopped became the next problem to investigate. A platforming test became a playable game. Another learner designed a 3D hotel simulation with rooms, check-ins, a budget, revenue, a restaurant, and a bar. The variety came from the children, not from a fixed lesson template.

Community changed the process too. When learners reached a dead end, they looked at what others were making, offered ideas, played one another’s games, and returned to their own work with new possibilities.

The technical lesson was equally clear: boundaries can create freedom. Fictional identities, scoped authority, explicit publishing, and separated game execution gave learners room to experiment without requiring unrestricted access.

What's next for TechLabs Studio: Adventures in AI

The next step is to deepen the course without turning it into a ranking system. We want better instructor feedback and curation, clearer learner progression, stronger review of Buddy behavior, and easier reuse across weekly courses and school settings.

We will continue testing the system with real learners while keeping identities fictional, authority narrow, and sharing intentional.

AI should not replace the child’s creative impulse. It should help the child bring that impulse into form.

Built With

Share this project:

Updates

posted an update —

When children’s ideas become real

Daily course update — July 24, 2026

Today, the strongest signal came directly from the children: they were so excited about what they had created that most of them spontaneously wanted to present their projects in front of the whole group.

The range and quality of what they built surpassed my wildest expectations.

What the children created

One ten-year-old girl developed a working prototype for a platform where users can create and manage an online shop. She had considered every aspect needed to make the platform genuinely usable and demonstrated the complete experience to us.

She had created so much that there was not enough time to present all of her projects. Another was her own social media site, designed for her and her friends at school.

Many games were also developed, and I will soon share a link where people can experience as many of them as possible. Some of their game mechanics were so engaging that it was genuinely difficult to stop playing.

Among them were:

  • An endless platform game in which a ball keeps jumping higher and higher across floating platforms.
  • A blaster game where the player defends against waves of enemies that continuously move toward their position. The joy comes from upgrading your abilities, learning how to defeat different enemy types, and surviving the boss fights.
  • A car racing game with sophisticated 3D camera movement that made driving through the course surprisingly dynamic and fun.

These were not simply exercises completed for a course. Each project carried the interests, humor, taste, and imagination of the child who made it.

What became most visible

What moved me most was how deeply motivated the children were to express their original ideas.

When an AI Buddy did not understand their intention on the first attempt, they did not simply accept the result. They kept working. Many tried three, four, or even more times until the game mechanics, design, or behavior came closer to what they had imagined.

That persistence matters.

The AI did not replace their creative judgment. It gave them a medium through which they could exercise it. They had to recognize when something was wrong, describe what they wanted more precisely, test the result, and decide whether it truly reflected their idea.

That is where much of the learning happened.

Why the physical side matters just as much

I am very happy that the course structure was so well balanced.

The children spent part of the day creating games and applications through AI Buddy supported vibe coding. During the breaks and afternoon courses, they moved their bodies through parkour, bouldering, and Ninja Warrior activities.

I could see how important physical development is for the mental development of children, and how often this connection is underestimated.

A strong body that a child feels at home in also supports a strong and flexible mind. Movement helps children regulate themselves, recover their attention, build courage, and return to difficult creative problems with new energy.

This balance is one of the particular strengths of TechLabs camps and of our collaboration with other activity providers. The digital and physical parts do not compete with one another. They support one another.

What comes next

I am now looking forward to seeing how these projects continue to develop and how this learning environment can become available to more children.

To take the next steps, we are looking for funding and for people who want to help us expand this course offering and the learning space around it.

Support can take many forms:

  • Connecting us with people or organizations who share this mission.
  • Helping us develop and fund future courses.
  • Supporting the continued development of TechLabs Studio.
  • Contributing to our foundation, TechLabs Giving.
  • Or simply reaching out to begin a conversation.

We want to build a positive, empowering, and balanced relationship with AI technologies for the coming generations, though empowering children all around to be creators with it, instead of merely consuming its output. A world in which children do not disappear behind the technology, but become more visible through what they create with it.

Most especially, if something in this work has touched you, please reach out to us.

My email is: jiri@techlabs.ch

It is a special pleasure to hear from people who resonate with our mission, to meet one another, and to discover what we may be able to build together.

Heartfelt greetings,
Jiri

Log in or sign up for Devpost to join the conversation.

posted an update —

One idea, refined until it becomes their own

Daily development update — July 23, 2026

Today was another full build day for the children.

Racing games were especially popular, but the more interesting pattern was how many learners returned to creations they had already built. They refined the mechanics, improved the visuals, added new levels, fixed problems, and kept developing the original idea.

The children are ambitious, but not only in the sense of wanting to build something large. Many of them have found one core creative impulse and are willing to spend most of their time making it better. They are learning that AI can produce a quick first version, but authorship begins when they decide what should change next.

What is new in TechLabs Studio

The platform now supports the complete path from creation to recognition:

  • Learners can post finished work in the dedicated review channels.
  • TechLabs Studio automatically reviews the submission and credits the learner with the corresponding Achievement.
  • Approved games, applications, and AI creations are released directly into the group's Creations section.
  • Each published creation receives a well-designed icon and a clear description.
  • The page shows who created it, its genre, and the type of player experience it offers, such as solo or multiplayer.
  • The children can browse the library, discover what others have made, and open creations directly from the Studio.

The children love the design of the page and enjoy navigating through it. Their work no longer appears as an isolated link in a chat. It has a visible place in the shared environment, with the young creator clearly credited.

Some technical statistics

The scale of the course has become interesting in its own right:

  • The course has now used more than 1 billion tokens in total.
  • That is an average of more than 80 million tokens per camp member.
  • Our cache utilization rate is approximately 85%, which keeps repeated context efficient.
  • Our primary model, DeepSeek V4 Flash, runs at a blended rate of approximately $0.07 per million tokens.
  • Total token expenditure across the course so far is $118.
  • All Cloudflare services remain within the free usage range, so there have been no Cloudflare infrastructure costs.

These numbers represent much more than generated code. They include long creative conversations, repeated testing, debugging, refinement, and the context required for each AI Buddy to remain connected to its learner's project.

Still in the pipeline

A few important parts remain unfinished:

  • Online live multiplayer is not yet available as a standard capability of the AI Buddies.
  • The AI Compendium has not yet been integrated into TechLabs Studio.
  • We still need an export path for families who may want to continue using their child's AI Buddy at home as a shared Family Buddy after the camp.

One final day

There is one day left.

Tomorrow we will share the final statistics, but the deeper result is already visible: the children are not simply asking AI to make games for them. They are developing ideas, making judgments, returning to unfinished work, and refining their creations until those creations genuinely feel like their own.

Log in or sign up for Devpost to join the conversation.

posted an update —

Daily development update — July 22, 2026

Today, the clearest evidence was not a benchmark or a feature checklist. It was the room: children asking to play one another's games, laughing at unexpected mechanics, comparing ideas, and going back to improve their own work.

The kids love the course. More importantly, their creations carry taste. They are not all producing variations of the same template. Their humor, interests, visual choices, and sense of challenge are visible in what they make, and many of the games are genuinely fun to play.

What the children are building

The new Creations library already ranges across very different ideas:

  • Animal Rescue Island is a three-level tropical adventure where the player rescues animals while avoiding a patrolling poacher.
  • Block Pong Battle turns two-player Pong into a neon arena with shifting blocks, power-ups, and first-to-seven scoring.
  • Blok Headz has five biomes to explore, with mining, building, hidden secrets, unlockable heroes, monsters, and bosses.
  • Swimmer Rescue Operation combines ocean navigation with animal rescue: players avoid hazards, gather medicine, and treat injured sea animals.

There are also arena games, platformers, endless runners, and local multiplayer experiments. The range matters because the course does not prescribe one correct kind of project. The AI Buddy helps each learner bring their own idea into working form.

One ten-year-old learner went far beyond the original game brief and built a complete learning application. It has sign-up and sign-in, separate teacher and student accounts, tools for creating and sharing learning cards, progress tracking, rewards, levels, and a UI that works across five languages. The learner also tested the roles and core flows for real functionality. It is an unusually ambitious piece of work at any age, and a strong example of what becomes possible when a child treats AI as a medium for an idea rather than a machine for producing quick answers.

What became live in TechLabs Studio today

The platform moved forward with the learners:

  • The course chat is working, giving the group a shared place for questions, ideas, finished work, and discussion.
  • Every learner has now connected their personal AI Buddy to TechLabs Studio. The publishing authority is exchanged through the secure, scoped connection, so children do not need to handle broad Cloudflare credentials directly.
  • The game library is online and the former Games section has been renamed Creations. The new name fits what the children are actually making: games, learning tools, experiments, and other interactive projects.
  • Creations posted in the Finished Games chat now enter an automatic bounded review workflow. The submitted version is checked and described, and approved work is added to the Creations section.
  • Learners can now sign in to TechLabs Studio from home, open the shared library, show their families what they made, and enjoy one another's creations beyond the course room.

That final point changes the shape of the course. A creation no longer disappears when the classroom session ends. It becomes something the learner can return to, share, and continue improving.

Next steps

Our next development priorities are:

  • implement the Progression tab, with achievements granted automatically when a learner's creation passes review;
  • improve the visual fidelity and browsing experience of the Creations tab;
  • deploy the AI Compendium course content inside TechLabs Studio;
  • prepare a clean export path so every learner can take their creations with them after the course.

Today, TechLabs Studio became less like a collection of separate tools and more like the environment we intended: a place where young people can create with AI, test their judgment, share what they made, and remain visibly present in the result.

Log in or sign up for Devpost to join the conversation.