Building a Little Prince 3D browser game with AI: one person, 16 days, full walkthroughDr. Xiaohui (case narrated)
The author is not in the games industry, and shipped a playable 3D browser game — The Little Prince: Across the Stars — in 16 days. This build log splits the work into seven stages: scoping, picking the tool chain, generating scenes and props, animating the character, fixing scale and camera feel, shipping, and a retrospective. Each stage names the agent / tool / MCP used, what broke, and how it was worked around.
Full walkthroughFull walkthrough7 stages
-
Step 1 of 7 Lock the scope: what fits in 16 days
The video keeps coming back to one admission: “I am not in the games industry”. That single fact rules out two things — following an industrial pipeline (level design → blockout → art → animation → audio → optimisation), and shipping anything large.<br /> So step one is not picking tools, it is squeezing the scope down to what one person can finish in 16 days: one small walkable scene, one protagonist, a few interactive props, one short path to walk end to end.<br /> This constraint keeps biting at every later stage — it decides how many scenes get generated, how far the character animation has to go, and which “properly required” steps can be dropped wholesale.Scope checklist for stage one Gotcha Start from “how many hours a day can I actually give this for 16 days”, not “what game do I want to make”. The second question reliably doubles your scope in week one.
-
Step 2 of 7 Pick the chain: Opus 5.5 + Tripo + Blender
The three-piece stack from the video, each doing exactly one job:<br /> · **Opus 5.5** — the thinking project manager: break the brief down, translate “the Little Prince watching the sunset on his planet” into something a generator understands, and propose a detour when you get stuck.<br /> · **Tripo** — the clay sculptor: one sentence or one reference image becomes an importable mesh, covering first drafts of scenes and props.<br /> · **Blender** — the finishing shop: fix what Tripo produced, rig and animate, render and export.<br /> Keeping the split explicit matters: do not expect one tool to cover the whole run. AI generation gets you from zero to “60%”; Blender gets you from 60% to shippable. That line runs through every stage below.How the three tools divide the work Gotcha Keep it at three tools on purpose. Every extra tool is another format conversion and another export-import round trip — the least affordable cost on a solo project.
-
Step 3 of 7 Generate scenes and props: one sentence into a mesh
This stage feeds the prompts written earlier into Tripo and batch-produces first-pass meshes for scenes and props. The video works **wide first, precise later**: generate several variants from the same description and pick the closest one, rather than tweaking the prompt repeatedly to converge on one result — the latter pays very poorly with generative tools.<br /> The output is a pile of `.glb` / `.obj` files that carry shape only: no correct scale, no materials, no animation. All of that is the next stage.Several prop variants generated in Tripo Gotcha Generate 3-5 variants and pick, rather than iterating the prompt. With generative tools the variance is more useful than the controllability.
-
Step 4 of 7 Making the Little Prince move: rigging and animation
A mesh does not walk by itself. This stage does two things in Blender:<br /> · **Rigging** — give the model a skeleton so that “raise the arm” has something to act on;<br /> · **Animation** — cut idle, walk and turn loops, good enough and no more.<br /> This is the step where **AI helps least**. Generative tools produce meshes, but getting them to emit usable skin weights is not realistic today — this part is hand work, and the biggest single time sink of the 16 days.Rigging and the walk cycle inside Blender Gotcha Do not over-build the animation set. Idle + walk + turn already make a character feel alive; the fourth clip onwards returns less each time.
-
Step 5 of 7 Polish: scale, style consistency, camera follow
The video says it plainly: AI made production faster, but it does not help with **integration**. Concretely, three things stick here:<br /> · **Character and scene scale disagree** — each looks fine alone, then breaks the moment they share a frame;<br /> · **Art style drifts** — props generated in different batches come out with different surface feel;<br /> · **Camera follow feels wrong** — the numbers are defensible, the game feel is not.<br /> There is no shortcut for these three; you iterate. The video's verdict: AI buys back the time from zero to something, not the time from something to playable.Before and after on scale and camera Gotcha Fix the unit at **generation time** (e.g. bake “1 unit = 1 metre” into the prompt) instead of rescaling later — far less work.
-
Step 6 of 7 Shipping: WebGPU is a hard requirement
The result lives at https://littleprince.dev/ and runs in a tab. The video calls out one thing specifically: **you need a browser that supports WebGPU**.<br /> Worth stating separately, because it is the hidden cost most easily missed on the “AI-generated 3D + publish to the web” route — however smooth generation is, it is worthless if the reader's browser cannot run it. Checking browser support before you ship is cheaper than patching compatibility afterwards.Checking WebGPU support in the browser Gotcha Put “requires WebGPU” somewhere unmissable on the page, not buried in the tech notes — a visitor who hits a blank screen will not go read the docs.
-
Step 7 of 7 Retrospective: will AI replace the games industry?
The last section lands softer than its title suggests: AI sped up **production**, but a game is still a product — publishing, distribution, monetisation do not disappear just because building got faster.<br /> The part that matters for the rest of us is the closing line: one person can now make the game they actually want to play. That was not true five years ago.<br /> The author says the game will be open-sourced and the full process published, so this build log will keep growing.The game page and the open-source plan Gotcha Treat “AI makes production faster” and “a product gets shipped” as two separate claims. The first payoff is certain; the second is not.