agentbull/ankole/pdf

PDF

The Agent Computer image installs Pandoc, two PDF engines, Poppler, QPDF, pdf-inspector (detect-pdf, pdf2md), and nano-pdf. This Skill owns how a PDF is read, made, checked, and corrected.

Important: If the human has already provided a brand palette or template, match that first. Otherwise, use design-md skills and use it as the design reference.

Read

Classify before extracting. detect-pdf samples content streams in milliseconds and reports whether a text layer exists at all:

detect-pdf in.pdf --json   # pdf_type, confidence, ocr_recommended, pages_needing_ocr

• text_based: extract locally with pdf2md in.pdf. It returns structured

Markdown with headings, lists, tables, and multi-column reading order. It owns PDF reading; do not read PDFs through pdftotext or markitdown.

• scanned or image_based: no text layer exists, and pdf2md fails with

a clear error. Route the file to the ocr Skill.

• mixed: extract the text pages with pdf2md in.pdf --select-pages 1,3-5,

then hand only the pages_needing_ocr list to the ocr Skill. OCR costs seconds per page; never send pages that extraction already covered.

When ocr_recommended is true, trust it over the type label. Treat low confidence as scan-suspect even for text_based: extract one page and check the output before trusting the whole file.

Create

Always produce the PDF with Pandoc from an authoritative source document. Do not lay out pages with a plotting library such as Matplotlib: it cannot reflow text, so it overlaps content without an error.

Pick the engine that fits the document. Both are installed, and each one takes a different styling language:

No LaTeX engine is installed. For citations use Pandoc's own --citeproc, which works with both engines. Do not install a TeX distribution inside a Job.

The image installs three text families, and design-md owns which one to use: IBM Plex Sans SC for text, IBMPlexMonoSCHalf for code, and 寒蝉锦书宋 (ChillJinshuSong) for serif. Sans and mono carry weights 300 to 700 including SemiBold 600; the serif has no SemiBold, so 600 resolves to Bold there. Fluent Emoji Color sits behind all three for emoji. Name the family explicitly, because an engine default can silently drop CJK glyphs. WeasyPrint accepts the ASCII alias ChillJinshuSong; Typst matches the family name inside the font file, so it needs 寒蝉锦书宋.

For HTML input, convert with pandoc --standalone --from markdown+raw_html --to html5 --css <stylesheet> first, then print that HTML through the same engine.

Check

A generated PDF needs a check, but not a visual one. Reading page images costs far more context than it returns and it forces conversation compaction. Run these instead:

pdfinfo report.pdf                       # page count, page size, embedded fonts
qpdf --check report.pdf                  # file structure
pdftotext -layout report.pdf out.txt     # text must agree with the source document
tr -cd '\f' < out.txt | wc -c            # form feeds must equal the page count
pdftotext -bbox-layout report.pdf -      # text box geometry

The geometry output carries every text box with its coordinates. Overlapping boxes mean overlapping content, and a box outside the page margin means clipped content. Both are the defects a visual pass looks for, and this finds them without an image.

Render a page image only for a page that the geometry check reports, look at that one page, and record what it shows.

Edit

nano-pdf corrects text in an existing PDF from a natural-language instruction. Use it for a typo, a title, or a date, not for layout work: rebuild the PDF from its source instead.

nano-pdf edit report.pdf 3 "Update the date from January to February 2026"

Page numbers can be 0-based or 1-based depending on the version. If the edit lands on the wrong page, retry with ±1. The tool calls an LLM, so it needs its own API key; check nano-pdf --help.