Image Processing

When to Use

• When the task requires describing an existing image, answering questions about it, or extracting information from it
• When the task requires deterministic manipulation of existing image files, such as downscaling, trimming, format conversion, or compression
• When the task requires traditional computer vision operations, such as filtering, edge detection, or object recognition
• When choosing between deterministic processing and AI generation or editing for an existing image

Image Processing Best Practices

• Decide by the deliverable: when the deliverable is an image a person will look at, prefer AI image generation or editing; when the deliverable is data extracted from an image, or the same pixels re-encoded, use code
• For image understanding tasks like description or Q&A, use the read tool to leverage native multimodal capabilities instead of writing code
• For generative image modifications that require creativity, such as semantic editing, style transfer, or upscaling/restoration/enhancement, use AI image generation and editing tools; MUST read the imagegen skill for route selection and prompt structure
• For basic image manipulation tasks like downscaling, cropping, format conversion, and compression, use Python with PIL/Pillow
• For traditional computer vision operations like filtering, edge detection, and object recognition, use OpenCV
• Reserve OpenCV for deterministic pixel-level or geometric pipelines whose result is data or a re-encoded file; for recognizing, classifying, or describing what an image contains, the read tool is more accurate and cheaper than hand-written detection code
• DO NOT use deterministic processing where new visual content must be created; deterministic processing only rearranges or re-encodes pixels that already exist

Route Boundaries

• When a resize or crop would change the aspect ratio, default to preserving all content unless the user explicitly accepts edge loss; follow the imagegen skill for that case
• DO NOT use PIL/Pillow, canvas, SVG, HTML, or scripted compositing to add or lay out text on a visual; text-bearing images are generated with their text already in place, as required by the imagegen skill
• DO NOT use OpenCV or other deterministic upscaling for resolution, clarity, or restoration requests; those are generative tasks owned by the imagegen skill
• DO NOT draw annotations, labels, arrows, callouts, dimension marks, or highlight boxes onto an image with PIL/Pillow, OpenCV, matplotlib, canvas, or SVG when the annotated image is itself the deliverable; annotate through AI image editing as required by the imagegen skill. Use code only when the user needs the annotation coordinates as data, or requires reproducible machine-generated overlays such as evaluation or dataset artifacts
• For architectural drawings, floor plans, interior layouts, product design views, engineering sketches, and similar design visuals, use AI image generation; DO NOT assemble them with plotting or vector-drawing code. Route to a deterministic workflow only when the user requires exact scale, verified measurements, or editable CAD/layout source files
• When the deliverable is a new visual, a chart, or a structured diagram rather than a transformation of an existing image, follow the routing rules in the imagegen skill instead of this one

Delivering Results

• Write transformed output to a new file path; MUST NOT overwrite a user-provided original unless the user asked for in-place replacement
• When a transformation can silently damage the result, such as unintended aspect ratio, rotation, color-space, or transparency loss, confirm the output once with the read tool; a single pass/fail check is sufficient, do not open an extended inspection loop
• Deliver the resulting image files, and keep the accompanying script only when the user needs to rerun or adjust the transformation