Building a Walking 3D Little Prince with AI: 16 Days, 115 Asset Bundles, and One Inverted-Controls Bug
For a Western developer evaluating AI-assisted game or 3D pipeline work, this is a concrete proof that a lone developer can assemble a multi-level walkable web game in just over two weeks, and the deep-dive of the actual build artifacts shows exactly which architectural choices made it feasible.
X user @Saccc_c shipped littleprince.dev, a 3D narrative game based on The Little Prince, claiming a 16-day build using Claude's Opus 5.5 for logic, Tripo for base 3D models, and Blender for polish. The game went viral on Hacker News and Chinese tech media, but the official site itself carries the most useful intelligence: it's a Unity 6 WebGPU build, and the embedded manifest lists 115 Addressables asset bundles totaling 2013.4 MB uncompressed.
The bundle structure reveals the game's architecture: one Addressables group per planet, with shared assets resident and high-res layers loaded on demand, which is what makes the scenes walkable instead of a slideshow. The pipeline the author describes is a clean, repeatable loop: split the book into independent level slices, generate base models with Tripo, refine in Blender and export glTF, assemble in Unity with per-planet Addressables groups, and drive script and scene wiring through Claude Code with a CLAUDE.md convention file.
The article also documents a concrete AI-workflow bug: a player reported inverted W/S and A/D controls after completing the watering quest, and neither the player nor the author could reproduce the trigger, a classic state-sequence-dependent failure. The missing open-source repo and the outstanding promise of a detailed walkthrough remain the boundary of what's verifiable.
Reading the site's HTML and asset manifest reveals more about the project's architecture and its AI workflow than the author's own narrative does, since the bundle prefixes and sizes expose the level-splitting and on-demand loading strategy directly.
The biggest asset bundle belongs to the geographer, a character with almost no plot; this illustrates that AI-assisted asset generation shifts the cost toward whatever textures or scenes the source text describes, not what the story actually needs.
The inverted-controls bug, which neither author nor player could reproduce, is a generalizable warning about AI-generated game code: long-horizon edits that depend on state sequences are the least testable and most likely to break invisibly.
The author's decision to pin conventions in CLAUDE.md before starting iterative edits is the most transferable and likely highest-ROI step in the entire pipeline.
The 16-day claim and the open-source promise remain unverified; what's checkable is the live site's build output and the reported tool division, so treating the rest as marketing is reasonable.