PDF

Work with PDFs through the bundled pdf.sh workflow. Treat structure extraction and visual rendering as complementary: parsed text is evidence about content, while rendered pages are the evidence for layout.

Hard requirements

• Preserve every input PDF. Write edits to a distinct output unless the user explicitly requests replacement.
• Use pdfplumber for text, table, image, and coordinate-aware inspection; use pypdf for page structure, metadata, page operations, and AcroForms; use ReportLab for new PDFs.
• Inspect an existing PDF before changing it. Render it first when layout, page order, form appearance, or visual fidelity matters.
• Create a new PDF from one executable Python builder. Patch and rerun the same builder rather than accumulating one-off scripts.
• Never assume successful text extraction proves that a PDF looks correct.
• Run audit, render every final page with Poppler, and inspect every page PNG at full size before delivery. A montage is only an overview.
• Fix clipped or overlapping content, missing glyphs, broken tables, incorrect page order, bad image crops, inconsistent page sizes, and wrong page numbers before delivery.
• Do not depend on Codex-private runtime paths or install Python packages globally.

Read the relevant references