Convert inputs to markdown

Source documents (bank statements, broker statements, receipts) come in mixed formats — PDF, xlsx, mhtml, html, sometimes plain text. Downstream skills work best on markdown: line-addressable, diff-friendly, easy to extract tables from.

This skill converts every file in regnskab{YYYY}/input/ (excluding the markdown/ subfolder) to a markdown file in regnskab{YYYY}/input/markdown/.

When to run

• After the user drops new files in input/.
• Before invoking build-workbook for the first time.
• When the user re-uploads a corrected statement (delete the old markdown first).

Tool

Microsoft markitdown is the converter. Install once:

pip install 'markitdown[all]' --break-system-packages

Naming pattern

The markdown filename keeps the original extension as a suffix. This avoids collisions when the same stem exists as both .pdf and .xlsx (a common case for broker statements):

broker2025.pdf  → input/markdown/broker2025__pdf.md
broker2025.xlsx → input/markdown/broker2025__xlsx.md

Use __{ext} (double underscore + extension) as the separator.

Per-file conversion

markitdown "regnskab{YYYY}/input/{filename}.pdf"  -o "regnskab{YYYY}/input/markdown/{filename}__pdf.md"
markitdown "regnskab{YYYY}/input/{filename}.xlsx" -o "regnskab{YYYY}/input/markdown/{filename}__xlsx.md"

PDF files become text. xlsx files become tables. mhtml / html become structured markdown. From this point on, downstream skills always read the markdown version — never the original.

OCR fallback

Plain markitdown extracts text but not text inside embedded images. Some broker statements render as scanned tables (the original PDF was a scanned document, not a born-digital one).

Trigger condition: a PDF input produces < 50 lines of markdown but the file is > 50 KB.

When triggered, enable the markitdown-ocr plugin with Mistral OCR v3 as the LLM client:

pip install markitdown-ocr --break-system-packages
markitdown-ocr "regnskab{YYYY}/input/{filename}.pdf" -o "regnskab{YYYY}/input/markdown/{filename}__pdf.md"

If Mistral isn't available, document the file as needing manual transcription in note{YYYY}.md.

Idempotency

Re-running this skill should be safe. The script:

1. Lists every file in input/ (top-level only — not the markdown/ subfolder).
2. Skips files that already have an up-to-date markdown counterpart (mtime(markdown) >= mtime(input)).
3. Re-converts files that are newer than their markdown counterpart.

Verification

After running, the user should see:

• One markdown file per source file.
• Each markdown file with > 5 non-empty lines (otherwise the conversion likely failed silently — flag it).

A simple check:

for f in regnskab{YYYY}/input/markdown/*.md; do
  lines=$(wc -l < "$f")
  if [ "$lines" -lt 5 ]; then
    echo "WARNING: $f has $lines lines"
  fi
done

What this skill does NOT do

• It does not interpret the content. That happens in build-workbook.
• It does not move or delete originals. The PDFs/xlsx stay in input/ for audit trail.
• It does not deduplicate. If two inputs have identical content but different names, both get converted.