PPT Image Content Recognition

Problem

A lot of critical information in PPT files is not inside editable text boxes. Ladder diagrams, circuit diagrams, equipment photos, and flow diagrams are often embedded as images, so plain text extraction only captures titles and short captions. For engineering notes, course materials, or technical decks, this easily leaves out wiring relationships, component labels, step sequences, and other structural details.

How It Works

The skill turns slides into inspectable page images on Windows and then asks a multimodal model to read the visual content. The main pipeline is:
- Slide export: Use WPS or PowerPoint COM to export each slide as PNG. This path works well with legacy .ppt files and can process multiple presentations.
- Image reading: A model with image input reads the exported pages and extracts visible text, relationships, and technical points. Verified models include qclaw/pool-minimax-m3 and qclaw/kimi2.6.
- Parallel processing: To keep the main session clean, the skill delegates batches to sub-agents, usually 10-20 images per batch, then merges the results.
- Output: The extracted content can be collected into a text report or written into a docx file with images and explanations.

Limits

It is best for image-heavy decks that still need human review before archiving. The skill is Windows-only because it relies on local COM automation. Recognition accuracy depends on the chosen multimodal model, and complex ladder or circuit diagrams may still miss wire numbers, pin labels, or small annotations. Very large files, especially over 500 slides, should be split into batches. If the default model does not support image input, switch to a vision-capable model explicitly.