Skip to content
English teacher · Seoul · product owner and designer

Classroom games that get a whole class talking.

I'm Jacobus Barnard. I design games and tools that run on one classroom projector, with no student devices, and I build them with AI coding agents. Five games, eight free tools and a teacher's classroom manager are live together as Class Cookies.

The Class Cookies home page: 'Pick a game. Put it on the screen.' with Games and Free tools buttons.
5classroom games live on classcookies.com
3,783automated tests passing across the six codebases
16,800words in 59 Seconds' 15 packs
694 / 869commits co-written with Claude Code

Counted on 7 October 2026 from the live sites and fresh copies of each production branch. How these were measured.

How I build

One person, many AI agents

I build Class Cookies and its five games by directing AI agents. I'm the product owner, designer and tester. Claude Code writes the code, nothing is built that I haven't decided on, and nothing is released until I say so.

Each of the six products, the five games and the Class Cookies platform, has its own repository and its own Claude Code sessions. Every session reads that repository's instruction file before it does anything else. Other sessions handle work that cuts across the products: this portfolio, and clean-up jobs such as checking every repository for a bug one session found in another.

Above the sessions sit two chat-based reviewers. Every session report goes to both. They check it against the brief, look for what was skipped or got through untested, and write the next prompt, which I paste into the session. They review the same work, and I pass each one the other's findings. As the brief puts it, "Disagreement is the product." Beside them, a separate chat-based Claude agent works on marketing: it analysed the market and drafted LinkedIn and Reddit posts.

I decide what gets built and design how it works in a classroom. I test the live sites myself and send back what I find. I use the games with my own classes most weeks, and colleagues have tried them in theirs. That is informal use, not a study, but it changes the design: Taco Trivia has no game-over screen, because final scores are counted after the last tile and a screen would cover the score panels.

On the production branches, 694 of the 869 commits carry a Claude co-author line. Git lists me as the author because I direct the agents and commit their work; it does not mean I typed the code.

The products themselves contain no AI: none of them calls an AI service.

  1. I write the brief: what to build, the rules, and a success criterion for each step.
  2. The product's session reads its instruction file, plans, and stops for my approval.
  3. It builds in small steps. A new test has to fail first. Done means typecheck, lint, tests and a production build all pass.
  4. The reviewers read its report and draft the next prompt.
  5. I decide and release, check the live site, and use the result in class.
Diagram of the workflow: I set the brief, the agent reads and plans, I approve the plan, the agent builds and tests, two chat agents cross-review, and I decide and release.
How the work is done. The quotations are from the instruction files and task briefs kept for the agents.
Writing

Reports

Correction, 16 October 2026. The report says a refund removed access. The only live purchase was my own test, and its refund did not remove access at the time. 59 Seconds never had paying customers. It also says "One teacher who can code": I direct AI agents to write the code; I don't write it by hand.

Jacobus Barnard
Contact

Jacobus Barnard

English teacher in Seoul. I use these games with my own classes most weeks, and colleagues have tried them in theirs: informal, everyday use rather than a formal study.