PDF To Markdown

PDF to Markdown

Create a validated Markdown sibling while preserving the PDF as the visual authority. Default to report.pdf plus report.md; do not retain parser JSON, coordinates, extracted assets, or chunks unless the user names a consumer for them.

Convert safely

1. Resolve each source PDF and intended output path. Process multiple PDFs one at a time so each result can be validated independently.

2. If the intended output exists, use --overwrite only when the user has explicitly authorized replacing that exact file. If that authority is missing, stop before conversion. Do not invent an alternate destination; use another path only when the user selects or approves it.

3. Check for the xberg executable. If it is missing, identify the current official installation method for the host platform and use applicable user approval, requesting it only when absent. Do not run an install script, package manager, or model download prewarming beyond that approval.

4. Resolve scripts/convert_pdf.py relative to this SKILL.md and run:

   python3 scripts/convert_pdf.py /absolute/path/report.pdf
   

   The wrapper writes report.md atomically, preserves the PDF, rejects empty or structurally inconsistent output, and never invokes a shell. It asks Xberg for JSON so it can validate every physical page, but publishes only Xberg's whole-document Markdown and retains no parser JSON.

   The wrapper's ordinary Xberg extraction uses these core flags:

   xberg extract /absolute/path/report.pdf \
     --no-config-discovery \
     --format json \
     --content-format markdown \
     --extract-pages true \
     --page-markers true
   

   When page markers are enabled, the wrapper also supplies a randomized private marker through --config-json; it later converts only those private markers to <!-- PAGE n -->. A direct fallback must use an equivalently collision-resistant private marker rather than public page comments during validation.

   The wrapper validates that Xberg's page records form a contiguous 1-based sequence and agree with its declared page count. If Xberg omits markers for pages it classifies as blank, the wrapper restores those markers from page metadata while preserving Xberg's Markdown body unchanged. It refuses publication when marker and page metadata cannot be reconciled. Do not rebuild the document by concatenating per-page Markdown because that can reset lists and damage structures spanning page boundaries. Do not replace the wrapper with a direct shell redirect when creating the final artifact.

5. For a PDF that contains scanned Korean pages, add --korean-ocr. This uses PaddleOCR with explicit Korean selection and OCRs only pages classified as scans while retaining native text elsewhere:

   python3 scripts/convert_pdf.py /absolute/path/report.pdf --korean-ocr
   

   The corresponding Xberg flags are:

   xberg extract /absolute/path/report.pdf \
     --no-config-discovery \
     --format json \
     --content-format markdown \
     --extract-pages true \
     --page-markers true \
     --ocr true \
     --ocr-backend paddle-ocr \
     --ocr-language korean \
     --ocr-scanned-pages
   

   The first PaddleOCR use may download models as part of the requested conversion. Tell the user before starting when network use or model storage is material in the current environment.

Do not OCR a born-digital PDF merely because its language is Korean. Use --force-ocr only when the entire document is image-only or its text layer is demonstrably broken; combine it with --korean-ocr for Korean documents.

If Python is unavailable but Xberg is present, reproduce the wrapper's exact flags with a command runner while preserving the same non-overwrite, temporary output, nonempty-result, and atomic-finalization guarantees. If those guarantees cannot be preserved, stop and report the missing capability.

Validate and escalate

After every conversion:

1. Inspect the beginning, middle, and end of the Markdown. Check headings, paragraph order, tables, lists, code, Hangul where expected, numbers, dates, missing pages, suspicious repetition, mojibake, and abrupt density changes.

2. Read the wrapper's validation summary. It reports physical, nonblank, and blank page counts and, when nonzero, how many blank-page markers it restored. Compare that total with an independent PDF page count when a PDF inspection capability is available. A blank page can legitimately contain no text, but it must still have a marker in the final Markdown unless --no-page-markers was requested. Repeated pdf_oxide dictionary-as-stream warnings are summarized rather than dumped; they mean malformed or unusual PDF objects were treated as empty streams, so inspect nearby visual content when fidelity is uncertain. The wrapper also reports Xberg's structured processing warnings with bounded output; treat them as page or pipeline-specific fidelity risks to inspect.

3. If ordinary extraction loses reading order, headings, tables, lists, or figure placement, rerun to a temporary comparison file with --layout:

   python3 scripts/convert_pdf.py /absolute/path/report.pdf \
     --output /absolute/path/report.layout.md \
     --layout
   

   This adds --layout --layout-strategy auto --use-layout-for-markdown to the ordinary Xberg command.

   Compare content fidelity, then replace the intended output only with the user's overwrite authority. Do not assume that the larger output is better. If layout extraction reports an ONNX Runtime API mismatch, treat that as the cause even if a later worker message says Mutex poisoned. On Windows, check whether an older C:\Windows\System32\onnxruntime.dll was loaded. Point ORT_DYLIB_PATH at a compatible runtime only with user authority; do not silently install or replace a system library. The wrapper publishes no partial layout output after this failure.

4. If Korean OCR still omits or corrupts content, retry with --force-ocr only when the native text layer is the cause. For complex visual pages that remain unreliable, report the failing pages and offer a Docling or document-VLM comparison rather than silently switching tools or installing another parser.

Finish

Finish only when the Markdown exists, is nonempty, passed representative source checks, and the source PDF remains unchanged. Report the PDF and Markdown paths, Xberg version, OCR or layout options used, physical/nonblank/blank page counts, restored marker count, summarized parser warnings, any overwritten file explicitly authorized by the user, and any pages or structures that remain uncertain.