PDF To English

PDF to English

Overview

Use this skill to produce English PDF copies that keep the original PDF's page size, images, tables, forms, and text locations. The normal output naming convention is the source filename plus _EN before .pdf, saved beside the source unless the user asks otherwise.

Do not install PDF/OCR/translation tools without explicit permission. First try local tooling already available in the environment.

Workflow

1. Identify the input PDFs and desired output directory/name. Default to <original_stem>_EN.pdf in the same directory.
2. Check available local tools. Prefer the bundled script scripts/pdf_layout_tools.py, which uses macOS Swift/PDFKit and requires no Python PDF packages.
3. Extract text with coordinates:

python3 <skill-dir>/scripts/pdf_layout_tools.py extract input.pdf --json > extracted.json

4. Render visual previews for layout inspection:

python3 <skill-dir>/scripts/pdf_layout_tools.py render input.pdf /tmp/input_preview --scale 1.4

5. Translate with LLM judgment. Preserve proper names, dates, measurements, symbols, status marks, addresses, construction grid references, and inspection terminology. Translate repeated headings consistently.
6. Create an overlay JSON. Use 1-based page numbers and PDF-point coordinates from extract. Each item should cover the original text with a white backing rectangle and draw the English text in the same area:

[
  {
    "page": 1,
    "x": 64,
    "y": 224,
    "width": 390,
    "height": 58,
    "text": "Inspector's comments: ...",
    "size": 6.6,
    "bold": false,
    "align": "left"
  }
]

7. Apply overlays:

python3 <skill-dir>/scripts/pdf_layout_tools.py overlay input.pdf output_EN.pdf overlays.json

8. Render the translated PDF and visually compare it to the original. Check that translated text is readable, stays inside intended boxes, does not cover photos/diagrams unintentionally, and preserves page count and page dimensions.

Layout Strategy

• For simple reports, overlay translated line blocks at the extracted Japanese text coordinates.
• For dense tables/forms, translate by visual region instead of every line when English text would be longer. Use smaller font sizes and concise phrasing.
• For photo reports, keep photos untouched and overlay only caption areas.
• For placeholders or template labels embedded in image areas, avoid blanking rectangles that cover photos or callouts.
• For vertical Japanese category labels, replace them with compact horizontal or multi-line English labels inside the same column.
• Keep copyright lines, logos, stamps, signatures, and status symbols unless the user explicitly asks to translate or remove them.

Script Reference

scripts/pdf_layout_tools.py has three subcommands:

• extract input.pdf [--json]: emits extractable text lines and bounds.
• render input.pdf output_prefix [--scale 1.4]: writes output_prefix_01.png, etc., for visual QA.
• overlay input.pdf output.pdf overlays.json: draws white-backed English overlay text onto the original PDF.

The overlay script is deterministic; translation, grouping, font-size choice, and visual QA remain LLM-suited tasks.

Validation

Before finishing, verify:

• Output PDFs exist at the requested location with _EN suffix when applicable.
• Page count matches the source PDF.
• Rendered previews show no unintended text/photo overlap.
• The final answer includes absolute links to the generated PDFs and notes any limitations, such as scanned text that required manual visual translation or text that was not extractable.

Source: metakirby5/.dots — distributed by TomeVault.