Image Tech Md Rebuilder

Claude Code Migration Notes

This skill was migrated from the local Codex skill registry for Claude Code. Use it as a Claude Code SKILL.md skill. Some source text may name Codex-only tools or channels; in Claude Code, use the closest available Claude Code tool, shell command, MCP/plugin integration, or local helper script. Keep secrets in Keychain, environment variables, or authenticated CLIs only.

Image Tech MD Rebuilder

Goal

Convert ordered screenshots of a non-copyable technical document into one coherent Markdown document. The deliverable must read like the original source document, not like an OCR dump.

Required Output Standard

• Produce a continuous technical .md file with the original document structure.
• Rebuild headings, paragraphs, numbered lists, tables, code blocks, and diagrams semantically.
• Use Mermaid for flowcharts, sequence diagrams, state flows, and architecture diagrams when the screenshot contains a diagram that can be faithfully represented.
• Preserve important screenshots as image references only when OCR/Markdown cannot represent the content reliably.
• Never deliver a final document organized as "screenshot 01 text, screenshot 02 text". That format is only allowed as an intermediate OCR scaffold.
• State the final absolute output path.

Workflow

1. Inventory images

   • Accept a single image path or an image folder.
   • Sort screenshots by filename and timestamp unless the user gives a different order.
   • Count images and inspect formats/sizes.

2. Run OCR into intermediate artifacts

   • Prefer the bundled script:

uv run --with rapidocr --with onnxruntime --with pillow python C:\Users\18357\.codex\skills\image-tech-md-rebuilder\scripts\ocr_image_folder.py --image-dir "<IMAGE_DIR>" --out-dir "<WORK_DIR>\ocr_output"

• If rapidocr fails, try rapidocr_onnxruntime, PaddleOCR, Tesseract, or a vision model if available.
• Keep raw OCR artifacts for audit: JSON, plain text dump, and scaffold Markdown.

3. Rebuild the document, do not summarize

   • Read the OCR dump in order.
   • Merge text that is split across screenshot boundaries.
   • Identify the source document hierarchy: title, intro, sections, Q&A headings, subheadings.
   • Reconstruct code from OCR into valid fenced code blocks.
   • Reconstruct tables as Markdown tables.
   • Reconstruct diagrams as Mermaid or compact ASCII if Mermaid would distort the content.
   • Remove screenshot artifacts: line numbers that are not part of code, duplicated repeated headers, OCR garbage, UI controls, cursor marks, and watermark clutter.

4. Correct OCR systematically

   • Fix common technical OCR errors:
     • FastAPl -> FastAPI
     • JsoN -> JSON
     • LLA / 1lm -> LLM / llm by context
     • MCPServer -> MCP Server
     • MCPClient -> MCP Client
     • StreamableHTTP -> Streamable HTTP
     • Server-SentEvents -> Server-Sent Events
     • 检素 -> 检索
     • 攻路 -> 攻略
     • Agont, Agint, Agein, Aget -> Agent
   • Correct obvious punctuation and numbering errors from OCR, such as 1。 to 1. in numbered lists.
   • Preserve domain terms, variable names, and code identifiers exactly when visible.
   • If a line is uncertain, compare against the original image before inventing missing content.

5. Validate before final

   • Confirm the final Markdown has no screenshot-section structure unless the original source was actually screenshot-indexed.
   • Confirm code fences are balanced.
   • Confirm image references, if any, resolve.
   • Search for obvious OCR garbage patterns: ????, 1111, 1I1I, uouua, Japeps, FastAPl, Agont, 检素.
   • Report residual uncertainty only if specific sections remain ambiguous after checking the image.

Output Naming

Use a clear final name near the image folder, for example:

• technical_document_rebuilt.md
• <project>_technical_document.md
• <source-folder-name>_rebuilt.md

Keep OCR intermediates in an ocr_output folder beside the final document.

Practical Notes

• Terminal display of Chinese can be mojibake on Windows. Verify file contents with Python Path.read_text(encoding="utf-8") rather than trusting Get-Content rendering.
• For code screenshots, prefer reconstructing likely valid code over preserving OCR line-by-line noise, but do not change the intended logic.
• For interview-prep documents, keep both "short answer" and "detailed answer" sections if present.
• For architecture/flow documents, Mermaid is usually better than embedding screenshots because it makes the result editable and searchable.