Exploded Atlas

Approach

Each subject is a hand-authored set of SVG <g> parts (6-9 per subject, ~22 total) arranged in their assembled position, with each part carrying a data-dir attribute (a precomputed unit vector for where it flies off to when exploded) and data-name/data-desc metadata. A single #explode range input drives a CSS variable consumed by each part's transform: translate(dir * value), so dragging it pulls every part outward along its own authored axis — cheap, deterministic, no physics needed. A second #rotate range input sets rotateX/rotateY with perspective on a shared .subject-stage wrapper, giving a pseudo-3D tilt over the flat SVG without ever touching a real WebGL context — more reliable to hand-build and gate within an hour than true 3D, and the pick's 'WebGL' framing is honestly replaced with CSS 3D transforms in the plan rather than claimed outright. Clicking a .part sets #label-name/#label-desc from its data attributes. Three tabs swap which subject group is .active (display toggle), with jet-engine active and already partially exploded by default so the effect is visible with zero interaction. Everything lives in one index.html with inline <style>/<script>, no fonts, libraries, images, or network calls.

The source post

https://x.com/arjunkshah21/status/2114732030533501057

Scoring

Pick

The Anatomy skill tweet is about a Claude/Cursor plugin that generates these illustrations on demand, which is a CLI/agent-skill, not a deployable artifact — but the underlying output (interactive 3D-style technical illustrations built from SVG + WebGL) is exactly the kind of thing that can be hand-authored once as a standalone, self-contained web gallery: a few subjects with real exploded/rotatable layers, built entirely from vector geometry and client-side WebGL with zero external calls at build or run time. It clicks in seconds (drag a layer, watch it separate; rotate; tap a part, see its label), needs no account, key, or dataset, and the interesting capability — real interactive 3D-feeling illustrations from flat SVG — is actually implemented rather than faked. Everything else in tonight's batch is either not a single web artifact (Real2Sem2Real's robot-demo pipeline, Haiku OS's bootable image, Nebius's CLI/sim-to-real framework), requires an account or backend to be interesting (Omnara's durable agents API), or can't honestly demonstrate its core claim as a static client-side build (MotionSites needs live design-to-animation interpretation of arbitrary sites; impeccable's 'eliminates AI slop' claim is a production-integration story that reduces to a plain style gallery here).

source: https://x.com/arjunkshah21/status/2114732030533501057

Review

mean 4.50/5

No change: the gate passed clean on the first attempt, NOTES.md reports no departures and nothing omitted for lack of a service, and nothing here recurs — history.json's last 30 nights show only two prior failures, both months-old wrangler/publish-step bugs already self-repaired, with no live pattern this night touches. The one soft spot (worth_it) traces to the pick stage choosing to hand-author a static instance of a 'generates on demand' tool rather than a weak execution of a good idea — that's a one-night judgment call on a thin batch (pick.json shows only one real candidate survived triage), not a repeatable rubric failure worth codifying yet.

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 needs no external service.