Fly an FPV drone through a desert canyon, raymarched in the browser

Exactly what the creator used

Claude Code

Any recent version

Node.js

Any recent version

HyperFrames

Any recent version

GSAP

Any recent version

WebGL2

Any recent version

puppeteer-core

Any recent version

FFmpeg

Any recent version

Claude Code — Claude Opus 5.5 (claude-opus-5-5) — wrote every file — your Claude plan Node.js — 24.x (24.13.1 here) — the HyperFrames CLI and the dev harness — free HyperFrames — 0.8.74 (npx hyperframes@0.8.74) — check, render (HTML → MP4) — free, runs locally GSAP — 3.14.2 (jsDelivr CDN) — the one paused timeline HyperFrames seeks — free WebGL2 — in Chrome (SwiftShader, software) — the raymarcher — free puppeteer-core — 24.22.3 (tools/ only) — dev snapshots and the clearance probe — free FFmpeg — any recent build — the Instagram re-encode, contact sheet and loop test — free Codex or any agent that can write files and run commands works too. No GPU is needed: HyperFrames renders WebGL in software, and this piece was built for that (see step 3). Provenance: Recorded September 28, 2026, and re-rendered September 29, 2026 after the rendering fix: Claude Code with Claude Opus 5.5 (claude-opus-5-5); GSAP 3.14.2 from jsDelivr; HyperFrames 0.8.74 rendering WebGL2 in Chrome 152 headless shell with SwiftShader at 1080×1920, 30 fps, 21.067 s (632 frames), no audio; the scene traced at 810×1440 and upsampled in the motion-blur pass. The render took 26 min 55 s with 4 workers on a 24-thread machine shared with other agents (beginframe capture, software GPU, HF_DE_STALL_MS=600000); the master is 38.3 MB, and the Reels copy was re-encoded with FFmpeg (libx264, CRF 21, aq-mode 3, maxrate 22M, faststart) to 19.3 MB. check passed with 0 errors and 4 warnings, all Chrome's "GPU stall due to ReadPixels" driver notes (about 11 minutes, with --timeout 180000). Camera (every frame): roll at most 14.6 degrees (mean 4.7 through the run), roll rate at most 30 degrees per second, no scripted rolls. Distance field (tools/lipschitz.js, 20,000 points per region): no rock inside any "safe" sphere in any region except a 0.003% residual of 0.19 m near the headwall base. Whole-video overstep audit (tools/audit.js, every third frame at 0.3 scale): 0.038% of pixels on average, none above 0.2% (before the fix: 0.18%, up to 1.0%). Loop test (Gaussian blur sigma 3, then SSIM, on the master): 0.8496 across the seam (last frame to frame 0) against 0.8519 and 0.8460 for the steps either side, and PSNR 24.42 dB against 24.69 and 24.14; the steps change smoothly on both sides (0.8636, 0.8593, 0.8554, 0.8519 into the seam; 0.8460, 0.8430, 0.8399, 0.8379 out of it). On the Reels copy: 0.8510 against 0.8536 and 0.8480. Numerically, the camera moves 0.2674, 0.2685 and 0.2699 m over the frames either side of and across the seam. Minimum clearance to rock along the flight: 1.53 m. No external images, footage, textures, models or generative-media models were used.