dream-loop

dream-loop makes AI coding agents better at visuals by giving them a picture to aim at. The agent first generates an image of what the finished scene should look like, builds toward it, then a separate critic agent compares a live screenshot against that target and sends notes. It loops until the critic is satisfied. MIT licensed, installed as a skill.

Install with npx skills add achimala/dream-loop or clone it into your skills folder. Your agent needs image generation, either built in like Codex or through an API key such as Gemini, plus vision input and ideally subagents. Blender is optional for custom 3D models, and the author prefers the Blender MCP over computer use.

Developers building browser games, 3D demos or visually heavy pages get the most from it. The showcase is a Three.js isometric fantasy scene from a one-hour, time-boxed prompt, with a live demo you can click around. The skill is free, but every loop spends image generation and model calls.

The catch is testing. The README says it has only been run with GPT-6 Astra in Codex, so on any other agent you are the test case. It is also a new repo from one developer.