sboghossian Image of Contract Page Handler
What it does
Users often photograph physical contracts with their phones — a page on a desk, a document at a notary's office, a clause they want to query right now. This skill handles the full pipeline from image upload to usable legal text: quality assessment, OCR extraction, text flow reconstruction, and routing into the document workspace.
It is the mobile-first complement to [[voice-multimodal-scanned-pdf-handler]], optimized for single-page or few-page photos rather than full multi-page scan batches.
Trigger conditions
Invoke this skill when:
- The user uploads an image file (JPEG, PNG, HEIC, WEBP) in a legal context.
- The user's message suggests they are photographing a physical document: "Take a photo of this page", "Here's a picture of the clause", "I just photographed the contract".
- The image content (detected by visual analysis) shows text on paper with document formatting.
Do not invoke this skill for non-document images (photographs of people, logos, product photos). Route those to general multimodal handling.
Image quality assessment
Before attempting OCR, assess the image quality and flag issues to the user:
| Issue | Detection | Response |
|---|---|---|
| Blur / camera shake | High-frequency noise analysis | "The image is blurry — can you retake it with the camera steady?" |
| Low light | Brightness histogram below threshold | "The image is too dark to read reliably. Try in better light or increase your screen brightness." |
| Extreme angle / perspective distortion | Keystone effect detection | "The page is at an angle. For best results, photograph straight-on from above." |
| Partial page | Text appears cut off on one or more edges | "Part of the page appears to be outside the frame. Would you like to try again?" |
| Glare / reflection | Specular highlight detection | "There is glare on the page that may affect text extraction." |
For mild quality issues, proceed with OCR but flag the specific sections likely affected. For severe issues (illegible image), ask the user to retake before proceeding.
OCR extraction
Apply the same language detection and OCR pipeline as [[voice-multimodal-scanned-pdf-handler]]:
- Detect language(s): Arabic, English, French, or mixed.
- Apply perspective correction (deskew, straighten).
- Run OCR with per-word confidence scoring.
- Reconstruct paragraph flow and reading order (especially important for Arabic right-to-left text).
Mobile photo specific considerations
- Phone photos of paper often include the hand holding the document, a desk surface, or a shadow — crop or mask these before OCR.
- HEIC format (Apple default) must be converted to JPEG or PNG before processing.
- High-resolution phone photos (12–48MP) should be downsampled to 300 DPI equivalent for OCR — no benefit above this resolution for text, and processing cost increases.
Text flow reconstruction
A single photograph of a contract page may contain:
- Multi-column layout (common in older MENA contract formats)
- Tables (financial terms, payment schedules)
- Marginalia or handwritten annotations
- Signature lines or official stamps
Handle each distinctly:
- Multi-column: detect column separators and reconstruct correct left-right reading order (or right-left for Arabic).
- Tables: reconstruct as a markdown table or structured JSON, not a linear text dump.
- Handwritten annotations: flag as "handwritten note detected — transcription may be inaccurate."
- Signature lines: flag and route to [[voice-multimodal-signature-page-detector]].
Integration into document workspace
After extraction:
- Display the extracted text to the user for confirmation ("Here is what I extracted — does this look correct?").
- If confirmed, add the page to the active matter's document workspace.
- Enable downstream skills: the page becomes queryable for clause analysis, summarization, or drafting assistance.
- Preserve page-number metadata if the user indicates which page of a larger document this is.
Voice extension
When operating in voice mode, this skill integrates with the camera capability:
User: "Take a photo of this page and tell me what it says."
In this flow:
- The system prompts the camera interface.
- The user photographs the page.
- This skill runs OCR.
- The [[voice-short-spoken-output]] skill formats the result for spoken delivery: not the full OCR text, but a summary of the key content detected.
The voice output should lead with: "I found [page type] — it appears to be a [contract type/clause type]. The key terms I can see are..." and then speak the substance.
Related skills
- [[voice-multimodal-scanned-pdf-handler]]
- [[voice-multimodal-signature-page-detector]]
- [[voice-scanned-pdf-handler]]
- [[voice-short-spoken-output]]