Visual UI Understander

Problem Context

When an engineer or agent receives a screenshot, the task is usually not simple image recognition. It may involve checking UI hierarchy, extracting copy, comparing design differences, and identifying accessibility risks. If a model answers freely, the output can be inconsistent and hard for downstream automation to parse.

How It Works

This skill is oriented toward AI Agent visual understanding. It covers screenshots, UI design review, OCR, multi-image comparison, and a11y checks. The core flow is: confirm the image path, choose the matching mode based on user intent, and return JSON or structured results. It supports general understanding, UI issue analysis, text extraction, multi-image input, and a short one-sentence summary for quick notification use cases. Internally it normalizes common image issues: resizing when the longest edge exceeds 2048px, converting RGBA/PNG/WebP to JPG, and reducing quality when the file exceeds 2MB. It retries network timeouts with exponential backoff up to 3 times, waits 2 seconds on rate limits, and returns immediate repair guidance for authorization or quota errors. Errors use the ERROR:type|description format so agents can parse them.

Boundaries and Notes

This skill depends on Tencent Hunyuan model service and the required cloud permissions. It requires TENCENTCLOUD_SECRET_ID and TENCENTCLOUD_SECRET_KEY. In multi-image mode, only the first image is visually analyzed, while the remaining images are passed as textual descriptions; for precise analysis of every image, call the skill separately.