WizWor
Build date:
What if choosing an old game felt less like searching a database and more like asking the strange wizard who lives inside the arcade cabinet?
WizWor is an agent-guided classic-game recommender wrapped in an 8-bit terminal. It interviews the player, learns their preferences, and turns that conversation into a recommendation from a local game catalog.
The experiment
I wanted to find the useful boundary between language-model judgment and deterministic software.
The agent gets to do the fuzzy human part: talk to you, notice what you care about, and translate vibes like “something weird but not punishing” into useful preferences. The recommendation machinery gets the boring, dependable part: work from a known catalog and return something the product can actually stand behind.
The interface leans all the way into the bit. CRT wizard. Synthesized speech. Chiptune audio. A little creature in the machine who would very much like to know what you played when you were twelve.
What worked
The newsletter
Get the build notes.
I write about building AI systems in public—what worked, what broke, and what I’d do differently.
The conversation makes preference discovery feel playful instead of form-like. More importantly, the agent does not need to own the entire system to make the experience feel agentic.
That became the useful lesson: give the model the part that benefits from judgment, and give ordinary software everything that benefits from certainty.
What I kept
WizWor became an early ratchet click for how I build agentic products: explicit tool contracts, bounded authority, deterministic seams, and an experience where the AI is visible because it is actually doing something worth seeing.