Study Image Reading
Output language: ALL learner-facing content MUST be Simplified Chinese.
Try native vision first
Use the available image-reading capability directly when possible.
- If you can see the image, produce the required mode output below.
- If image reading fails or the model has no vision, use the external vision API script.
One failed native attempt per session is enough evidence; do not retry every image.
External vision API
python3 ~/.claude/skills/study-img/scripts/recognize.py <image> --mode <mode>
First-use configuration: ask for provider type, base URL/API key, and vision model. Store config in ~/.config/study-img/config.json, chmod 600, and never repeat the full API key in conversation.
Modes
| Scenario |
Mode |
Required output |
| Scanned textbook page / photographed paper / handout |
--mode ocr |
Structured Markdown transcription; formulas as LaTeX; figures as [图：...] placeholders with enough detail to locate them. |
| Textbook/courseware figure, coordinate plot, table image, flowchart, chart |
--mode figure |
Teaching-grade description complete enough to redraw or convert into a lecture figure/table. Include axes, labels, variables, trends, data rows, and the conclusion. |
| Learner handwritten answers |
--mode answer |
Verbatim transcription; preserve errors; LaTeX formulas; use 【?】 for illegible characters. |
| Unsure |
no mode |
Comprehensive recognition. |
Workflow hookups
- Scanned PDFs: render flagged pages with
extract_pdf.py --render-scanned, recognize, and merge into internal/textbook/chapter-XX.md.
- Image-heavy PPT slides: export with
extract_pptx.py --render-images, recognize, and merge into internal/textbook/chapter-XX.md.
- Lecture figures: when a
[图：...], chart, curve, or table is important for understanding, recognize it with --mode figure; then study-teach must include the useful visual/table/formula in the lecture JSON with source_ref.
- Handwritten answer grading: transcribe with
--mode answer, show uncertain parts to the learner, then hand to study-quiz for grading.
Caveats
Vision output can misread formulas and numbers. Cross-check against surrounding text, dimensions, and internal consistency before teaching or grading from it. If a figure/table remains doubtful, say so and ask the learner to confirm from the original.
1---2name: study-img3description: Image-reading sub-skill (orchestrated by study-assistant; also usable standalone). Use for ANY study material that must be visually inspected: scanned textbook pages, courseware figures/charts/diagrams, photographed exam papers, photos of handwritten answers or notes, and png/jpg/jpeg/gif/webp/bmp files. Supports OCR, teaching-grade figure descriptions, and verbatim handwritten-answer transcription.4---56# Study Image Reading78**Output language: ALL learner-facing content MUST be Simplified Chinese.**910## Try native vision first1112Use the available image-reading capability directly when possible.1314- If you can see the image, produce the required mode output below.15- If image reading fails or the model has no vision, use the external vision API script.1617One failed native attempt per session is enough evidence; do not retry every image.1819## External vision API2021```bash22python3 ~/.claude/skills/study-img/scripts/recognize.py <image> --mode <mode>23```2425First-use configuration: ask for provider type, base URL/API key, and vision model. Store config in `~/.config/study-img/config.json`, `chmod 600`, and never repeat the full API key in conversation.2627## Modes2829| Scenario | Mode | Required output |30|---|---|---|31| Scanned textbook page / photographed paper / handout | `--mode ocr` | Structured Markdown transcription; formulas as LaTeX; figures as `[图：...]` placeholders with enough detail to locate them. |32| Textbook/courseware figure, coordinate plot, table image, flowchart, chart | `--mode figure` | Teaching-grade description complete enough to redraw or convert into a lecture figure/table. Include axes, labels, variables, trends, data rows, and the conclusion. |33| Learner handwritten answers | `--mode answer` | Verbatim transcription; preserve errors; LaTeX formulas; use `【?】` for illegible characters. |34| Unsure | no mode | Comprehensive recognition. |3536## Workflow hookups3738- Scanned PDFs: render flagged pages with `extract_pdf.py --render-scanned`, recognize, and merge into `internal/textbook/chapter-XX.md`.39- Image-heavy PPT slides: export with `extract_pptx.py --render-images`, recognize, and merge into `internal/textbook/chapter-XX.md`.40- Lecture figures: when a `[图：...]`, chart, curve, or table is important for understanding, recognize it with `--mode figure`; then study-teach must include the useful visual/table/formula in the lecture JSON with `source_ref`.41- Handwritten answer grading: transcribe with `--mode answer`, show uncertain parts to the learner, then hand to study-quiz for grading.4243## Caveats4445Vision output can misread formulas and numbers. Cross-check against surrounding text, dimensions, and internal consistency before teaching or grading from it. If a figure/table remains doubtful, say so and ask the learner to confirm from the original.
2362094903-ops