vision-inspection-workflow

name	vision-inspection-workflow
description	Read screenshots, OCR text, recognize image content, and inspect video-derived frames with the LM Studio multimodal model instead of relying on markdown embeds alone.

Vision Inspection Workflow

When To Use

• User wants text read from a screenshot, photo, scan, UI, chart, or other local image.
• User wants image recognition, visual question answering, or captioning on a local file.
• User wants to inspect video content and the workflow can be reduced to representative frames.
• An image was embedded into chat for display, but the model still needs actual pixel analysis.

Workflow

1. If the target is on screen, capture it first with as_screenshot_capture or locate the existing image file.
2. If the goal is exact text, numbers, equations, or UI readouts, call as_vision_ocr on the image source with a precise request about what to decode.
3. Start with the default fast path on the vision tools. Only retry with fast=false if the fast answer is incomplete, clearly wrong, or too uncertain for the task.
4. If the goal is to click, drag, crop, or otherwise act on a specific button, slider, icon, or label, call as_vision_target to get approximate image-relative coordinates and bounds for the requested targets.
5. If the goal is a detailed question about inner elements, interface state, or a specific region, call as_vision_focus with a concrete question.
6. If the goal is broader recognition or interpretation, call as_vision_recognize with a concrete question and expected output.
7. If the image is hard to read, use a tighter crop, a higher-resolution source, or another screenshot before retrying. Prefer a better crop before escalating from fast=true to fast=false when the problem is local readability.
8. If the target is a video, inspect it with as_media_probe, then use as_media_transform to extract representative frames and run the vision tools on those frames.
9. Use as_file_embed only when you want the image rendered in the transcript for the user; do not rely on the embed itself as proof the model has seen the pixels.
10. Report uncertainty when text is tiny, rotated, low-contrast, or partially obscured.

Preferred Tools

• as_screenshot_capture
• as_vision_ocr
• as_vision_target
• as_vision_focus
• as_vision_recognize
• as_media_probe
• as_media_transform
• as_image_identify
• as_file_embed

Output

• source image or frame paths
• OCR text or recognition answer
• uncertainty notes
• follow-up frame or crop suggestions when needed