dream-loop Gives Coding Agents a Gated Critic Loop for AAA 3D Visuals

A developer named Achim Alexander has published dream-loop, an open-source agent skill that gives any capable coding agent a repeatable process for producing high-fidelity 3D visuals. The skill, authored by Anshu Chimala, wraps Blender, an image generator, and a subagent critic into a closed feedback loop where the agent keeps rebuilding a scene until a separate judge model says the live screenshot matches a AAA-quality target render.

Coding agents are decent at writing Three.js or Blender scripts, but left alone they tend to output flat, blocky, obviously procedural scenes. Dream-loop reframes the problem as an image-matching task with a strict scoring rubric, and lets the agent iterate against that until it hits a threshold.

How the loop actually runs

The skill is essentially a structured prompt spec (a SKILL.md file plus install glue) rather than a model or a framework. The process has five stages: the agent dreams up a high-quality target screenshot using image generation, builds toward that target, hands the live screenshot to a separate AI critic that compares it to the target and gives feedback, loops back to building until the critic is satisfied, and optionally dreams up an even better target based on the current state.

The critic is the interesting part. Judging is meant to be done by a fresh subagent with a clean context each round to keep it objective and cheap, and it receives the latest live screenshot, the concept image, and from round two on the previous round's screenshot and verdict. The judge scores on a gated 0-10 ladder:

• Tier 1 (0-3), shape: camera framing, composition, and rough position and scale of every major object.