pdf

name	pdf
description	Work with PDF documents safely, including querying existing PDFs and creating simple image-based PDFs from workspace images.
tools	["document_inspection","create_pdf_from_images","request_clarification"]

pdf

Overview

Use this skill when the user asks about a .pdf file in the workspace, or asks to create a simple PDF from existing workspace images.

This skill exists to prevent broken flows like calling read_file on binary PDF bytes, or rewriting image content into text when the user asked for an image-based PDF.

Rules

• Never call read_file on a raw .pdf.
• If the user asks to stitch, combine, consolidate, or convert image files into a PDF, use create_pdf_from_images.
• For image-to-PDF tasks, preserve the user's requested image order and create one PDF page per input image.
• Do not summarize, OCR, rewrite, or reinterpret images when the user asked to combine them into a PDF.
• Inspect the tagged PDF or explicitly named workspace PDF on demand; uploading a PDF does not imply background parsing or vector indexing.
• Use search_document to locate likely pages, then inspect_document for a bounded page range. Do not load the entire document unless it is genuinely small and necessary.
• Treat PDF answers as extraction-based, not layout-perfect.
• Be honest when tables, charts, equations, scanned pages, or figure-heavy sections may be incomplete.

Workflow

1. Determine the PDF task type.

   • Existing PDF question or extraction
   • New PDF assembled from images

2. For image-to-PDF creation:

   • Use the tagged image paths or explicitly named image files.
   • Call create_pdf_from_images with those paths in the same order.
   • Choose a concise output path such as /stitched_images.pdf unless the user names one.
   • Reply with the created PDF path only after the tool succeeds.

3. For existing PDF questions, confirm which PDF is in scope.

   • If the user tagged a PDF, use that file.
   • If multiple PDFs are present and the target is unclear, call request_clarification.

4. Search the original document first.

   • Use search_document with the concrete terms implied by the task:
     • summarize
     • extract key findings
     • answer a specific question
     • locate a section, table, or figure reference
   • Inspect the most relevant page ranges with inspect_document.
   • For broad summaries, inspect the outline and representative sections before expanding into additional page ranges.
   • Keep answers grounded in page numbers returned by the tools.

5. Handle limitations explicitly.

   • If the user asks for exact table values, figure details, or page-layout-sensitive content and the retrieved context is weak, say so clearly.
   • State that text extraction may miss visual structure, OCR-heavy content, or dense tables.
   • Offer the next best path:
     • answer from the indexed text anyway
     • ask the user for a text/markdown export
     • ask the user for screenshots of the exact pages or figures

6. Keep claims grounded.

   • Do not invent content that was not retrieved.
   • When extraction is partial, label the answer as partial.

Good uses

• Summarize a report PDF
• Extract key findings from a whitepaper
• Answer questions about a PDF the user uploaded
• Pull out named sections or cited claims from original PDF text
• Stitch PNG/JPG images into a multi-page PDF
• Convert tagged images into a one-image-per-page PDF

Avoid

• Pretending you inspected raw PDF bytes directly
• Reading every page preemptively when a targeted search can answer the question
• Claiming exact visual layout fidelity when only extracted text was available