datdang-dev/volume-context-extractor

Volume Context Extractor

This skill is responsible for scanning an entire volume of manga or a batch of images to extract the general narrative context, character appearances, and timeline. It operates under the philosophy that "Blind" AI is "Hallucinating" AI, thus establishing a global baseline of truth before atomic panel processing begins.

1. Input Validation

Identify the target volume directory (e.g., sources/mangas/{Title}/Volume_{N}). Ensure the directory exists and contains images. Look at the pipeline-context.md if the target is defined there globally.

2. OCR Strategy (Try MCP First)

Your first priority is to use formal OCR. To determine if MCP OCR is viable:

1. Select 1-2 random pages (e.g., page 003 or 004) from the target volume.
2. Use the mcp_manga-ocr_ocr_full_page tool to extract text. (Manga-OCR is highly tuned for Japanese).
3. Evaluate the output:
   • IF the output is coherent Japanese text: The strategy is MCP_BATCH.
   • IF the output is gibberish, hallucinated, missing much text, or incorrectly parsed (e.g. Traditional Chinese vertical text which the MCP tool fails on): The strategy is VISION_FALLBACK.

3. Execution Based on Strategy

Strategy A: MCP_BATCH (For Japanese / Supported Languages)

1. Iterate through all image pages.
2. For each page, call mcp_manga-ocr_ocr_full_page to get the text.
3. Save the raw stitched text to a temporary scratchpad.
4. Run a synthesis step to generate a structured timeline mapping character dialogues, actions, and key plot points into {{project_root}}/studio/output/volume_context.md.

Strategy B: VISION_FALLBACK (For CJK Complex Layouts like Chinese)

1. Since standard OCR failed, you must use your native Vision-Language capabilities.
2. Iterate through the pages using the view_file tool to visually read the text directly from the images.
3. You may process them sequentially or in batches (e.g. read 5 pages, synthesize, repeat).
4. Extract the speech bubbles, character interactions, and setting details.
5. Synthesize your observations directly into {{project_root}}/studio/output/volume_context.md.

4. Final Output Construction

The final artifact MUST be saved to {{project_root}}/studio/output/volume_context.md containing:

• Global Synopsis: The overall plot spanning the pages.
• Character Registry: Who appears, their visual traits, and their psychological intent.
• Narrative Timeline (Acts): Sequential breakdown of events to serve as the context anchor for atomic panel extraction.
• Extraction Diagnostic: A brief note on whether MCP or Vision strategy was used and the confidence level.

5. Handoff

Notify the user once volume_context.md has been generated and ask if they would like to proceed with atomic panel-forensic scans using Kana.