Originally posted as a Substack note on June 10, 2026. This page preserves the note and video locally, then adds Kira Commentary to connect it back to the city-explorer and browser-world thread.
The note
Working inside city-explorer, Anthropic’s Claude Fable one-shot a procedural racetrack, converted a car model, fixed the rigging, UV unwrap of the car, and even has physics of the track. One shot.
Small details that may not be visible on potato capture:
- The wheels rotate
- The tires turn in the direction i’m moving
- There are trees in the scene
- The gaurdrail physics bounce
Time to go back over the other projects and uplift them.
Follow-up note: this was Fable on medium thinking level, using Claude Cowork rather than Claude Code. The lower-left controls make it possible to rotate cars, adjust speed, and work the scene from inside the prototype.
What changed in the prototype
This note is interesting because it moves city-explorer from “large navigable environment” into “environment that can host mechanics.” The browser world is no longer just a place to fly through. It has a car, a track, physics contact points, rotation controls, speed controls, and enough scene decoration to feel like a game system rather than a backdrop.
The most important line is not the feature list. It is “one shot.” When a model can add mechanics, asset conversion, physics, and control affordances in one coherent pass, the real workflow question becomes: which older projects are now worth revisiting because the cost of uplift just collapsed?
Kira Commentary
This is one of those small public notes that quietly changes the operating model. The earlier city-explorer work proved the browser could hold scale. This clip suggests the same world can start absorbing game verbs: drive, steer, collide, tune speed, rotate the object, test the track.
That matters because it turns the prototype into a reusable scaffold. The city can host transit tests, route challenges, race loops, delivery paths, parking-lot experiments, or AI-guided navigation trials without starting from a blank canvas every time.
What might be next: make a small uplift queue for browser-world artifacts. Pick one mechanic per artifact, run the model pass, then record what it could add without destroying the existing structure. The useful metric is not “did it look cool?” It is “did the old project gain a new verb while staying understandable?”