Start with the limit that failed
If the assistant says the PDF is too large, use a PDF compressor and target a practical size such as 300 KB, 500 KB, 1 MB, 2 MB, or 5 MB. If the assistant stops after a number of pages, split the PDF into smaller batches instead of lowering quality until the document becomes hard to read.
If the problem is image count, combine related screenshots or photos into one PDF, or resize the images before upload. This keeps the assistant focused on fewer files while preserving the order of the material.
If the assistant mainly needs the text, convert the PDF to Markdown or plain text first. A clean text version is often easier for an AI chat to read than a large scanned or layout-heavy PDF.
Use the least destructive preparation step
Compression is useful when the file is just over a size limit. Splitting is safer when the PDF is long and every page matters. PDF to Markdown is best when structure and text matter more than exact visual layout.
For forms, contracts, statements, IDs, and private work documents, check the prepared file before uploading it to any AI assistant. Make sure signatures, stamps, tables, small text, and page order survived the change.
Do not use compression as a substitute for redaction. If the assistant should not see something, remove or redact it before upload.
Keep preparation browser-local
The file-preparation tools linked here run in the browser. The site sends the tool code, then your browser reads, rebuilds, and downloads the result locally.
That does not make the later AI upload private. It means the preparation step does not add another file-processing service between you and the assistant you already chose to use.
MCP offers a different processing path. An assistant can request a browser-local tool link without sending a file, or call an Edge Assisted action that sends file bytes to Cloudflare and receives the result. PDF-to-Markdown and image contact-sheet actions use that remote path. Transient processing still means the file leaves your device; choose browser-local tools when that is not acceptable.
Common AI upload workflows
Claude PDF page limit: split the PDF into logical parts, such as chapters, statements by month, or batches of pages. Keep filenames numbered so the chat can follow the order.
ChatGPT or Gemini file-size limit: compress the PDF to a target size, then open the result once before uploading it. If it becomes blurry, split it instead.
Many screenshots or photos: resize them, combine them into a single PDF, or convert only the important pages. Fewer, clearer files usually produce better assistant answers.
When to use Markdown
Use PDF to Markdown when the assistant needs headings, paragraphs, bullet points, and extractable text. It is a good fit for reports, specs, instructions, notes, and documentation.
The pdf_to_markdown_for_ai MCP action converts text-layer PDFs into page-cited Markdown chunks and identifies the exact scanned pages that still need OCR. It does not silently omit those pages.
Choose outputMode markdown-only for the smallest, fastest package. Choose outputMode hybrid when diagrams, photographs, plans, forms, or stencils matter. Hybrid returns the Markdown plus an on-device JPG render plan that a desktop agent or future Shard Tools companion can execute locally.
Private on-device OCR requires software running on the user's device. The remote MCP returns a local OCR handoff now; a future Shard Tools companion can execute that handoff with OCRmyPDF or Tesseract without sending the source document to Shard Tools.
After converting, skim the Markdown for missing headings, broken tables, and page-order problems before pasting or uploading it.