Visual inspection and review (gemini-vision)
Inspect rendered interfaces, mockups, and wireframes with a multimodal vision model. Use when auditing visual hierarchy, confirming UI bug fixes, reading complex diagrams, or transcribing wireframes. For pixel or color analysis of camera or capture images, read the image directly or script it instead.
Workflow
Capture or locate image:
- Web application: capture a full-page or element screenshot with a headless-browser or screenshot tool.
- Local mockup / wireframe / diagram: use the local file path.
Inspect and label:
node <skill-dir>/review.mjs <image-path>To use a custom prompt, model, or OpenAI-compatible endpoint:
node <skill-dir>/review.mjs <image-path> --prompt "Label all buttons, inputs, and form sections in this mockup" --model "gemini-2.5-flash"review.mjsneedsVISION_API_KEY,OPENAI_API_KEY, orGEMINI_API_KEYin the environment (or in.envor.config/vision/envunder the home directory).VISION_BASE_URLandVISION_MODELare optional. If no key is set, stop and ask the user to provide one. Do not invent review results. Set variables for one run like this:POSIX: GEMINI_API_KEY=<key> node <skill-dir>/review.mjs <image-path> Windows (PowerShell): $env:GEMINI_API_KEY='<key>'; node <skill-dir>/review.mjs <image-path>Verify contrast deterministically: Vision models approximate; never trust raw color numbers from a model. If the inspection flags potential contrast issues, test the hex pair directly:
node <skill-dir>/wcag.mjs "#FOREGROUND" "#BACKGROUND"Iterate: Make code changes, re-capture the screenshot, and run a second review to confirm resolution. Cap review cycles at two passes. If the second pass still fails, report what remains to the user.
Diagram and wireframe transcription
When analyzing photographed whiteboard wireframes or rough sketches:
- Ask specifically for component bounds, text transcriptions, and visual hierarchy.
- Identify primary focal elements before inspecting secondary navigation.
- Flag any ambiguous gestures or handwritten text for human confirmation.