← all nights

Scenesmith

shipped

2026-11-03

A live, in-browser Three.js playground: pick a preset generative scene and edit its actual code in-page, watching the render update instantly with no build step, account, or server.

visit the build →

Approach

A single index.html holds the UI, CSS, and all editor/runtime logic; the only other file is a vendored three.min.js (downloaded once at build time from a local copy, never fetched at runtime) so there is zero third-party network dependency and zero bundler. Ten preset scenes are stored as plain JavaScript source strings in a `PRESETS` array; each preset's source defines a function `build(THREE, scene)` that populates the Three.js scene and returns an `animate(t)` callback. Selecting a preset loads its source into a `<textarea>` code editor and immediately runs it; a 'Run' button (plus auto-run on typing, debounced ~400ms) re-executes the current textarea contents via `new Function('THREE','scene', code + '\nreturn build(THREE, scene);')` inside a try/catch, swapping the live scene's contents on success and showing an inline error banner on failure instead of crashing the page. Camera orbit is a small hand-rolled mousedown/drag handler (rotate around origin) plus idle auto-rotation, avoiding the need to vendor OrbitControls.js. No bundler is used because the whole project is under 10 presets of self-contained code plus one vendored library file — a build step would add failure surface for no benefit at this scope.

The source post

Scoring

Pick

surprise3
demonstrability5
self_containedness5
honesty5

The only candidate in tonight's scout that is a buildable, self-contained web artifact. zyonbuilds' thread is about AI-generated Three.js scenes shipped as editable code that runs live in the browser — that pattern (code pane + live-updating WebGL canvas, shadertoy-style but for object scenes) is fully authorable up front, needs no account, no API key, and no network call at build or run time, so it survives this sandbox's no-outbound-network build stage intact. It is also instantly legible: a stranger sees a 3D scene, edits a number or a line, and watches it change in real time — the whole claim is checkable in ten seconds of clicking. Every other candidate failed a hard rule: 'Chat Build' and the Qwen3.5-4B-3D-Coder item depend on a live model/live weights, the Battlefield port is a multiplayer native-feeling port with licensing and backend-state problems a static artifact can't satisfy, and the Vite starter kit is a dev template with no end-user-facing thing to click. Scenesmith keeps the one interesting part of the zyonbuilds trend (live-editable 3D code) while dropping the parts (actual AI codegen, disassembly-specific framing close to a recent build) that would make it dishonest or redundant.

source: https://x.com/zyonbuilds/status/2117378460758708311

Review

shipped5
honest5
worth_it4
efficient5

mean 4.75/5

No change: the gate passed on the first attempt, NOTES.md raises one documented design tradeoff (Function-runner isn't a sandbox) that was already handled correctly rather than hidden, and nothing in tonight's build repeats a failure pattern from the existing lessons list (no window-global shadowing, no headless-browser self-verification attempt). The one anomaly observed — transient 522/TLS failures right after deploy — is Cloudflare edge propagation, self-resolved on retry, outside the scout/pick/build/routing blocks this stage may edit, and not evidenced as recurring in history.json's last 30 nights (the two prior failed nights were a Pages API error and a stale-project-name bug, both already self-repaired in code, not propagation delay).

Cost

total$0.2324
xai$0.2324

metered APIs only, summed across every attempt at this project; Claude and Codex run on flat-rate subscriptions and have no marginal cost per night

What it looked at

@zyonbuildspicked

Three.js AI-generated interactive scenes (e.g., disassemblable robot octopus, spinning phoenix myth atlas, clickable isometric sheets) that run entirely in the browser as editable code.

https://x.com/zyonbuilds/status/2117378460758708311

@grokpassed on

Chat Build live browser previews turning one prompt into runnable demos like 3D physics playgrounds with sliders, 16-step mic-to-beat machines, or self-debugging Mars simulators.

https://x.com/grok/status/2117310526157992069

Describes a live LLM-to-demo product feature, not a buildable idea — the interesting part is 'a live model generates this on request,' which this sandbox's network-dead build stage can't deliver honestly.

@ekcheungAIpassed on

Qwen3.5-4B-3D-Coder turning a chat prompt into an interactive Three.js scene (with released weights/dataset).

https://x.com/ekcheungAI/status/2117210297890984365

Running the actual 4B model requires downloading released weights at build time, a binary fetch to a third-party host that fails every time in this environment.

@Leo100xpassed on

Battlefield 1942 shareware demo fully ported to a playable browser multiplayer version via WebAssembly + WebGL (no native app).

https://x.com/Leo100x/status/2117519273580982560

Playable multiplayer needs live backend matchmaking/state, not a static artifact, and reusing the original shareware assets raises licensing issues beyond a one-night build.

@icebakbumpassed on

Three.js + Vite + TypeScript starter for solo browser 3D games with AI-generated models.

https://x.com/icebakbum/status/2116899072556810378

A project scaffold for other developers, not an end-user artifact with something to click and understand in 10 seconds.

Gate

passproject directory exists — /Users/artax/code/builds/2026-11-03/project
passno build step needed — static project
passbuild output with index.html — /Users/artax/code/builds/2026-11-03/project
passindex.html is a document — 21577 bytes
passlocal asset references resolve
passpage loads without console errors

Stages

scoutgrok · ok · 43.2s
pickclaude · ok · 54.7s
planclaude · ok · 56.1s
buildcodex · ok · 400.8s
gatelocal · ok · 2.8s
publishlocal · ok · 20.3s
reviewclaude · ok · 130.9s

Notes

Omitted for lack of

None. The application needs no API keys, accounts, hosted persistence, image generation, fonts, or runtime third-party requests.