Scrollcraft

Approach

Ship one self-contained index.html with inline CSS and vanilla JS — no build step, because six CSS/canvas scenes and a scroll-position mapper are well within plain HTML/CSS/JS and a bundler would only add failure surface for a one-night build. Six full-viewport <section class="scene"> elements sit in a scroll container; an IntersectionObserver plus a scroll listener compute each scene's progress (0-1) and write it to a CSS custom property (--p) that drives transforms (rotateX/Y, translateZ, clip-path, canvas particle positions) — so the effects are pure function-of-scroll-position, not timers. A fixed sidebar lists all six scene titles as dots (so the full scope is visible without scrolling), and clicking a dot calls scrollIntoView to jump straight to that scene. Scene 1 (the cube) also runs a slow idle CSS keyframe rotation on load so something is visibly alive before the visitor scrolls at all. The particle-field scene uses a plain 2D <canvas> (no WebGL) seeded with ~80 particles whose positions are recomputed from scroll progress on each animation frame. Everything — markup, styles, scene logic, and the six captions — lives in the one file; there is no data to fetch and nothing that can 404.

The source post

https://x.com/himanshubuildss/status/2108232309996650863

Scoring

Pick

Every other candidate tonight failed a hard rule before taste even mattered: Coderblock and the 300+ agent repo and elements.dev are link collections or multi-agent builder tools, not something we can ship as our own deployable artifact; open-code-review is explicitly a CLI; ArtCraft is a suite of MCP-controlled apps, not one artifact; Beam's weights haven't dropped yet and Step 5 requires downloading a model we can't fetch in this sandbox. Scrolltide is the one candidate pointing at something genuinely buildable as a static, self-contained page — scroll-driven 3D/CSS effects are plain code, no API, no login, no dataset. We're not lifting their components; the brief is to build an original showcase in that style, authored tonight, so Honesty and Self-containedness both score top marks. Surprise is the honest weak point — scroll-triggered 3D sites are a known web-dev genre, not an AI-specific novelty — but it's the only candidate that clears every rejection rule, so it wins by default on a thin field.

source: https://x.com/himanshubuildss/status/2108232309996650863

Review

mean 4.25/5

No change. The pipeline itself ran cleanly — single-attempt gate pass, no repairs, low cost — so there's nothing in run.json or NOTES.md to fix mechanically. The real weakness tonight is pick quality on a thin scouting night (7 of 8 candidates were structurally unbuildable: link directories, a CLI, unreleased weights), which pick.json and the published README already name plainly. That's a real signal worth watching, but it's evidence from one night, not the three-night pattern the revert rule asks for, and history.json doesn't carry prior nights' scout/pick detail to check for recurrence. Tightening scout.md's query on a single data point risks narrowing it off good candidates for the wrong reason. Worth revisiting if the next two nights show the same 'thin field, hard-rule-filtered down to one weak survivor' shape.

Cost

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

Gate

Stages

Notes

Omitted for lack of

None. This implementation requires no services, accounts, API keys, or runtime network requests.