Low Card Wild

An original multiplayer card game taken from concept to written rulebook to a functional cross-platform application in thirty-two working days — by one person directing AI agents, in languages he had never written a line of.

One round, unedited — solo against two AI opponents. 27 seconds.
32Working days
507CI tests
1Human

A game was the point, not the product. It is a bounded problem with fixed rules and success conditions you can decide rather than argue — a concrete task for a frontier tool instead of a nebulous one. I wrote the specification and made the structural calls; the agents wrote the code, in Dart and TypeScript. The server adjudicates every match, so the rules had to hold up twice, in two languages, against implementations trying to drift apart. I ran it as a program: I was the TPM, the agents were the engineering staff — planning, implementation, review, and release gates, all staffed by models, all decided by me.

Lessons learned

Directing the agents

  1. Match the model to the work. Token limits forced a tiering: Opus plans, Haiku writes, Sonnet reviews. Arrived at to stay under budget; kept because it fit — the expensive model on the decisions, the cheap one on the typing.
  2. Getting stuck is where the money goes. A bug neither I nor the model understood burned tokens faster than anything else. What fixed that wasn't tooling — I got better at prompting, at reading the code's behavior, at knowing when to escalate to a bigger model, and at handing over screenshots instead of describing what I saw.
  3. Self-hosting has a lot of potential but isn't there yet. Tried local models on a MacBook and a 70B on rented GPU. Quality, tooling, or economics failed every time; a frontier subscription beat all of it on cost and overhead. The open stack is roughly where Linux was in the early 2000s — and a new model or coding tool released tomorrow could change that instantly.
  4. The skill was the point. I'm better at directing agents than I was four months ago — not so I can do the work myself, but so I can lead and evaluate the people who do, and speak about it straight.

Beyond agentic coding

  1. Best practices matter even more when AI is doing the coding. There were SDLC practices I'd never personally implemented before this. Experience in and around engineering taught me they had to be followed — especially with AI writing the code: an automated test suite managed by CI, and code hygiene enforced through PRs and reviews.
  2. The rewrites weren't losses. Language and toolkit choices had to be revisited as I learned more and as limitations in earlier choices manifested — JSX to Swift to Flutter; a fragile Firebase multiplayer solution replaced with industry-standard Nakama. Necessary components of rapid prototyping and pushing the boundaries of my own experience.
  3. Some limitations I had to experience before I knew they existed. I moved to automated testing on physical devices when the simulators started thermal-throttling my MacBook — slowing the tests and everything else running beside them.

This project pushed the limits of my AI experience while calling on decades of domain knowledge and troubleshooting — and gaining new experience with technologies I had yet to encounter.