pdf-editing

PDF Editing

Tool Route

Use the bundled tools; do not install nano-pdf, pypdf, qpdf, Poppler, or another PDF package during a task.

• Call read_files with one paths item for a normal PDF; extraction is prepared in the same call.
• For scanned or image-only pages, call pdf_render only on needed pages and read the returned images with the current model. If it cannot read images, ask for readable text or a vision-capable model; do not install a recognition engine. Mark unreadable fields instead of guessing.
• Use create_pdf with source_type:"markdown" for a text-first new PDF or source_type:"html" for tables and custom layout.
• Use edit_pdf for deterministic edits to an existing PDF: merge, extract, delete, reorder, rotate, watermark, visible text/image overlay, and form filling.
• Use pdf_render on every changed page and representative unchanged pages before delivery.
• Every edit_pdf output is a new path. Read that new output with read_files;

its first call prepares extraction before returning content.

Editing Rules

1. Work from the user's source but always write edit_pdf output to a separate .pdf; never overwrite the only source.
2. Page numbers passed to edit_pdf and pdf_render are 1-based. Inspect page count/content before destructive page operations.
3. Use extract_pages to split or select pages. Use merge with input_paths in the exact desired order.
4. Treat overlay_text as visible annotation/replacement artwork only. It does not remove underlying text and is not secure redaction.
5. Arbitrary replacement of existing PDF text is not structurally reliable. Prefer editing the source DOCX/XLSX/PPTX and regenerating the PDF; otherwise use a disclosed visual overlay and inspect it.
6. Do not claim legal-grade redaction, digital signing, encryption, or password removal: these are not part of the current built-in contract.
7. For forms, use exact field names. Render the result and confirm values, checkboxes, and flattening behavior in a PDF viewer when fidelity matters.

Delivery Check

Check page count and order, rotations, changed-page renders, clipping, watermark placement, form values, extracted or visually read text, output path, and source preservation. For a watermark or overlay, explicitly confirm from the rendered pages that it is legible, positioned as intended, and not clipped. Return the final PDF path, untouched source path, actions performed, pages affected, and any remaining viewer-specific review.