ui-judge

Automate UI visual evaluation by capturing URL screenshots at multiple viewports and scoring them against a five-criterion rubric. A two-step LLM process generates a structured JSON verdict and identifies impacted selectors for targeted fixes.

Visual regression checking compares UI screenshots against expected quality standards. This skill uses a five-criterion rubric evaluated through a two-call LLM process per screenshot to produce a structured verdict detailing visual quality and affected selectors.

Capture screenshots at different viewport sizes for UI testing by triggering an automated evaluation that runs per viewport. Multi-viewport runs generate separate judge-<viewport>.json files, while single-viewport runs write results to judge.json.

Yes, UI evaluation with Claude vision identifies specific CSS selectors to fix. The two-step LLM process generates a structured JSON verdict and then isolates impacted selectors for targeted UI corrections.

Yes, you can use Playwright screenshots for LLM-based visual quality scoring. The skill captures URL screenshots and applies a two-call LLM process to evaluate them against a five-criterion visual quality rubric, outputting a JSON verdict.

Automated UI evaluation is advisory only and will not fail your build, as the workflow always exits with code 0. It provides a structured verdict to guide manual fixes rather than enforcing strict pass/fail gates.