# Improve a generated 3D scene with a screenshot and critic agent loop

PaperChase agent use case (free feed entry), version 1.0. Source: [@SKatalystAI (Sinda) on X](https://x.com/SKatalystAI/status/2111809289442435225), 2026-10-18.

## What they did

One prompt produced a first version, then three rounds of screenshot, critique and revise took it from what the creator rated 2 out of 10 to a moody night scene.

## Worth knowing

The key detail is that the critic is a separate, fresh agent, so it judges the image instead of defending its own code. The same loop works for UI, game art and landing pages. Ratings are the creator's own.

## Workflow template

```text
Apply this workflow to [your project]:
1. Have the agent build the first version.
2. Have it take a screenshot of the running result.
3. Start a fresh agent, show it only the screenshot and the goal, and ask for a blunt critique.
4. Feed the critique back to the builder and repeat two or three times.
Stop and ask me before anything irreversible: deploys, payments, sends or deletes.
```

## Steps

1. **Have the agent build the first version.** Have the agent build the first version.
2. **Have it take a screenshot of the running result.** Have it take a screenshot of the running result.
3. **Start a fresh agent, show it only the screenshot and the goal, and ask.** Start a fresh agent, show it only the screenshot and the goal, and ask for a blunt critique.
4. **Feed the critique back to the builder and repeat two or three times.** Feed the critique back to the builder and repeat two or three times.

## Credit

Workflow reported by @SKatalystAI (Sinda) on X: https://x.com/SKatalystAI/status/2111809289442435225

Summary and template written by PaperChaseWebb in our own words. Questions: mgmt@paperchasewebb.com