Ship Harness Bench
A benchmark to evaluate the quality of different LLM coding harnesses. Same model, same prompt, different tools.
9 agents · same prompt · different results
A benchmark to evaluate the quality of different LLM coding harnesses. Same model, same prompt, different tools.
"Create a beautiful, relaxing, 3D ship simulator for the browser with the sun in the horizon alternating day and night and when you click on the sea the water makes a wave"
The model is important, but the harness is a worthy optimization. Same model, vastly different results depending on the tooling.
A good browser-use tool is one of the most useful skills. Seeing, interacting with, and verifying output in a real browser is a game changer.
Memory and other tools could also be useful. Context retention and auxiliary capabilities shape output quality as much as the model itself.
Import maps, modular Three.js, day/night cycle with wave controls.
Self-contained single HTML with inline Three.js bundle. Keyboard sailing controls.
Clean vanilla JS with verification scripts. Minimal and focused approach.
Enhanced variant with bun-built bundle. Polished visual effects.
Second iteration with refined aesthetics. Static JS, no build step.
Vendored Three.js, minimal vanilla implementation with orbit controls.
Vite-built with custom fonts, time-of-day slider, and ambient sound.
Vendored Three.js with OrbitControls, straightforward implementation.
TypeScript + Vite. Seagulls, splashes, wave audio, ambient soundscapes.