Rules for agents (read first)

• Parse once to a file — never re-parse the same PDF to search again.
• Default search = bounded grep: grep -n -i -C2 "term" file | head. Get context in one command, batch lookups, cap with head.
• Use the query skill only when grep would flood — the discriminating term is common/ambiguous AND the corpus is too large to scan (many matches to sift). query returns a bounded, ranked top-k (small -k), keeping context small. Add --language <lang> for non-English.
• Tables where column alignment must survive -> pdf-to-text (this skill preserves Markdown tables fine; prefer pdf-to-text only when whitespace/columnar alignment matters). Don't read the source PDF as an image to get its text (vision is the fallback only for scanned/image-only PDFs with no text layer).

PDF to Markdown

Convert PDFs into structured, semantic Markdown that preserves the document's logical structure — headings, tables, lists, and reading order — rather than producing flat text. This is significantly higher quality than reading a PDF directly with the read tool, which only extracts raw text without structure.

Related Nutrient skills

This is one of a family of Nutrient document skills that install separately but are built to work together (same underlying binary). When your task calls for one, add it the way your agent installs skills:

• query — to search the converted file instead of reading a large output back into context: ranked BM-25 search that returns only the top line windows. The recommended way to pull facts from big conversions ("parse once, query many").
• pdf-to-text — for plain text instead of Markdown, when column/tabular alignment must survive or the consumer can't parse Markdown.

Usage