← all nights

One-Shader Photo Studio

shipped

2026-11-02

A full photo studio — key light, reflections, soft shadows, a backdrop — rendered entirely inside a single WebGL fragment shader, no textures, no libraries, no assets.

visit the build →

Approach

A single full-viewport <canvas> gets a WebGL2 context and draws one fullscreen triangle whose fragment shader raymarches a signed-distance-field scene: an infinity-cove backdrop (floor curving into a back wall), a cylindrical pedestal, and a hero object (sphere, torus, or rounded capsule, picked by a uniform int so no new geometry or assets are needed to vary it). Lighting is a single raymarched key light with IQ-style soft shadows (penumbra from marching toward the light and tracking the minimum ratio of distance to height), a cheap ambient term from the SDF normal's upward component standing in for occlusion, and a fixed dim fill/rim light so unlit sides aren't pure black. The floor reflection is a second, shorter raymarch along the mirrored ray, blended by a Fresnel term so it fades toward the horizon like a real seamless-paper floor. Camera orbit comes from mouse-drag updating azimuth/elevation uniforms; two sliders move the key light's azimuth and height; one button cycles the hero shape; two checkboxes toggle the reflection and shadow terms off so the viewer can see what each one is contributing. No bundler or library is used because the entire deliverable is one shader string plus a few dozen lines of uniform-wiring JS — a build step would add a toolchain for code that fits in one file and must run instantly.

The source post

Scoring

Pick

surprise4
demonstrability5
self_containedness5
honesty5

It's the only candidate that is both a real web artifact and actually self-contained: the whole claim is a raymarched scene computed per-pixel in one GLSL shader, so there's nothing to fake and nothing to fetch — it either renders a convincing studio or it doesn't, and you see which within a second of loading the page. The other four either aren't deployable single web artifacts (an autonomous-driving stack, an agent framework with GitHub/ticket integration) or are 'meta' tools whose interesting behavior is composing/previewing live agent interactions — exactly the shape this sandbox's offline build can't support honestly, since the real payoff would depend on a live call or get faked with fixtures.

source: https://x.com/redsyntx/status/2116564647528567004

Review

shipped5
honest5
worth_it4
efficient5

mean 4.75/5

No change: the gate passed on the first attempt, NOTES.md reports no departures from the plan and nothing omitted for lack of a service, and the shader checks out as honestly wired on direct read rather than just asserted. The one thing worth watching rather than acting on: run.json again logs 'pick -> claude: only candidate', the third-plus occurrence of this routing, but tonight's four rejections (Autotoken, Gofannon, DeepSeek dsh-superagent-prototype, Autoware 2.0) cite reasons the 2026-10-29 scout.md fix doesn't target — not verifiable in a static demo, backend/CLI shape, robotics framework with native components — rather than the paid-API-key/weights-download/native-app trio that fix excluded. That's plausibly the same underlying 'scout surfaces non-static-web-artifact candidates' problem wearing a different hat, but it's one night of this specific shape, not the three nights of evidence the revert rule asks for before touching scout.md again.

Cost

total$0.2059
xai$0.2059

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

@hexapodepassed on

Autotoken** (open-source hackathon experiment by @hexapode): converts repeated agent tasks into reusable programs with a live self-contained web demo. Interesting because it eliminates redundant agent work via on-the-fly program synthesis.

https://x.com/hexapode/status/2116390868878053521

The interesting part — synthesizing a reusable program from repeated live agent tasks — isn't something a stranger can verify in a static demo without either a real agent backend or faked traces, and it's not a 10-second-legible artifact either way.

@DanKornaspassed on

Gofannon** (open-source Apache 2.0 toolkit + web app by @DanKornas): guided interface for composing agent tools/data/paths then previewing real-time interactions in lightweight browser UIs. Interesting as a pure web-based prototyping environment.

https://x.com/DanKornas/status/2116658014404960350

A guided interface for composing agent tools is only honestly interesting once it's actually invoking tools/data in real time, which this sandbox's network-free build can't do without faking the 'live' part.

@redsyntxpicked

One-shader photo studio** (ray-tracer written by Opus 5.5, posted by @redsyntx): full photo studio (lighting, reflections, etc.) implemented in a single WebGL shader with zero external libraries or assets. Interesting as a tiny, instantly runnable browser demo of AI-assisted graphics.

https://x.com/redsyntx/status/2116564647528567004

@ANNANEXUSLABSpassed on

DeepSeek dsh-superagent-prototype** (MIT-licensed TypeScript repo released 31 Oct): event-driven autonomous agent system with persistent tickets, GitHub integration, and recovery. Interesting as a small, self-contained research codebase explorable in-browser or via simple static hosting.

https://x.com/ANNANEXUSLABS/status/2117041225060544877

An event-driven agent system with persistent tickets and GitHub integration is a backend/CLI system, not a single deployable, headless-browser-smoke-testable web artifact.

@AutowareFdnpassed on

Autoware 2.0** (major open-source autonomous-driving release): end-to-end AI + hybrid stacks with new model-management and package-composition tools. Interesting for its concrete, browser-viewable docs and examples of composable AI driving components.

https://x.com/AutowareFdn/status/2117171093098340553

A full autonomous-driving stack is a robotics framework with docs and native components, not something that reduces to one deployable web page.

Gate

passproject directory exists — /Users/artax/code/builds/2026-11-02/project
passno build step needed — static project
passbuild output with index.html — /Users/artax/code/builds/2026-11-02/project
passindex.html is a document — 10968 bytes
passlocal asset references resolve
passpage loads without console errors

Stages

scoutgrok · ok · 41.6s
pickclaude · ok · 35.7s
planclaude · ok · 54.1s
buildcodex · ok · 138.8s
gatelocal · ok · 2.9s
publishlocal · ok · 17.4s
reviewclaude · ok · 163.1s

Notes

Omitted for lack of

None. The studio requires only browser-native WebGL2; no external service is needed.