Inogen Lab
Comparing model performance on a one-shot task
We gave four Claude models an identical brief: build a cloth and rope physics toy as one self-contained HTML file, with next to no follow-up. A fifth version was produced with Sonnet (a smaller model) but this time working from a detailed written specification rather than the bare brief, demonstrating the power of a more guided approach.
Play with each one, then guess who made what: Haiku, Sonnet, Opus, Fable, or Sonnet with the spec.
Try tearing, pinning, and cutting the cloth...
Note that clunky controls on mobile is as the models wrote it. Pressing the full screen button is recommended for mobile.
Read the exact briefThe brief
Build a cloth and rope physics toy as a single self-contained HTML file (inline CSS and JS, no external libraries). Requirements: Simulate cloth as a grid of particles connected by constraints, using Verlet integration. Include a hanging cloth (pinned along its top edge) as the default scene. The user can interact with the mouse: drag to pull the cloth around, and hold right-click (or a toggleable "cut mode") to slice through constraints, tearing the cloth. Constraints should also break automatically when over-stretched, so the cloth can tear under its own weight if pulled hard. Include at least one other scene or object type, such as a free-hanging rope or a cloth pinned at only two corners. A scene switcher should let the user move between them. Provide controls for gravity, wind, and stiffness (constraint iterations), adjustable live. A reset button restores the current scene. The simulation must stay stable under abuse: fast mouse movements, extreme slider values, and heavy tearing should not cause it to explode or freeze. Aim for smooth performance at 60fps with a cloth of at least 40x30 particles. Render on a canvas. Make it look good and feel satisfying to play with - visual style, colours, and any extra flourishes are up to you. Show a brief hint on screen so a first-time user knows how to interact without instructions.
Mouse recommended. Each specimen runs exactly as its model wrote it; if one misbehaves, that is part of the result. Restart reloads it.
Who made what?
Each column matches exactly one specimen. Tap a cell to place your guess; placing a maker twice moves it. Fill the grid, then reveal.