Pro workflow

Agent skills · A check on the result · MIT

What it does

The loop with a judge: build a first pass, take a screenshot, give the screenshot and the target to a fresh sub-agent that scores composition 0 to 3, lighting 0 to 3, materials 0 to 3 and details 0 to 1 and lists every gap as a fixable instruction; fix them all, repeat. Stop at 8 of 10 with an acceptable frame rate. If the best score has not risen a full point in two rounds, stop tweaking and rethink the whole approach once; if that fails too, ask the user. The judge's prompt is written out in full and is the part worth lifting.

Get it

Loaded by the skill for 'high tier' users or when the user says Pro. The judge needs: the latest screenshot, the target image, last round's screenshot and verdict, and the prompt in this file.

Source on GitHub
Needs

An agent that can start sub-agents with a clean context and can see images

License

MIT

License notes are on the pack's page

From

dream-loop

Everything from dream-loop in the gallery

Back to the Asset Library