Uploading Documents Properly

Uploading a document is not a prompt. It is only context. The quality of the result depends on whether you tell ChatGPT what the file is, what you need from it, and what kind of output would actually help.

Show a file upload paired with a clear task label: summarize, extract, compare, audit, or transform.

The core idea

Treat a document upload like a working source. Name it, define its role, and specify the transformation or extraction you want. ChatGPT is much more useful when the file has a job: compare, summarize, extract, audit, rewrite, or classify.

The reason this matters is that document uploads create a subtle trap. The file feels like it speaks for itself. You drop a PDF into the conversation and assume the model knows what you need. But the model does not know whether you uploaded a contract to check for liability clauses, to summarize it for a colleague, or to compare it against a second document you plan to upload next. Without a stated job, the model defaults to a generic summary, which is almost never the thing you needed most.

It also helps to choose the right output shape. A memo needs different output from a checklist, a table of extracted fields, or a list of contradictions inside the document. Think of the output as the form factor for the next step in your workflow. If you plan to forward the result to a colleague, a clean summary works. If you plan to review it yourself against the original, a structured extraction with page references works better. If you need to decide whether to act on the document, a findings-plus-caveats format works best.

How it works

1. Before uploading, decide whether the whole document matters or only a section.
2. Tell ChatGPT what the document is and what success looks like.
3. Request an output that makes it easy to spot omissions or errors, especially for dense documents.
4. When working with long documents, consider whether pasting a relevant excerpt would give you a faster, more focused result than uploading the entire file.

What skilled users do differently

A novice uploads a file and types a short question, treating the upload itself as the prompt. They expect ChatGPT to figure out what matters. The result is a broad summary that covers everything equally, which means it emphasizes nothing useful.

A skilled user does something different before they even attach the file. They decide what role the document plays in the task. Is it the thing being reviewed? The reference standard? The source of data? They name that role explicitly, then define the extraction or transformation they need. They also specify the output shape, because they know the answer needs to fit into whatever comes next: an email, a decision meeting, or a deeper analysis pass.

Skilled users also know when not to upload. If the relevant content is two paragraphs on page fourteen, they paste those paragraphs and save the model from processing eighty pages of irrelevant material. The upload feature is powerful, but it is not always the most efficient path to a useful answer.

File upload limits (as of March 2026)

• 512 MB per file hard limit
• Images: 20 MB max; Spreadsheets: 50 MB max
• Text and documents: up to 2 million tokens per file
• Up to 20 files per message (increased from 10 in February 2026)
• Free tier: 3 files per day; Plus: approximately 80 files per 3 hours; Pro: unlimited
• Unsupported formats: video, audio, executables, password-protected documents

As an alternative to manual upload, ChatGPT supports Google Drive and OneDrive integration for connecting files directly from cloud storage.

Two worked examples

Example 1: a vague upload

Summarize this PDF.

This prompt is weak because it gives the document no role. ChatGPT does not know whether you need a high-level overview, a clause-by-clause breakdown, a comparison against another document, or an extraction of specific fields. The result will be a generic summary that covers everything equally, which usually means it is useful for nothing specifically. You will read it, realize it missed the part you cared about, and then ask a follow-up question that you should have asked first.

Example 2: a focused extraction

I am uploading a policy document.

Task: extract the parts that affect onboarding and access control.
Output format:
1. a 6-bullet summary of the sections that matter most
2. a table of concrete obligations, owners, and deadlines if they exist
3. a short list of ambiguities or missing details I should review in the original document

Do not summarize irrelevant sections just to be complete.

This version defines the document's role (a policy being reviewed for onboarding relevance), the extraction target (obligations, owners, deadlines), and a verification aid (ambiguities to check manually). The instruction to skip irrelevant sections prevents the model from padding the answer with material you do not need.

Example 3: a comparison task

I am uploading two versions of a vendor agreement (Q3 draft and Q4 final).

Task: identify every substantive change between the two versions.
Ignore formatting, numbering, and cosmetic edits.

Output format:
- a table with columns: section, Q3 language (brief), Q4 language (brief), significance (low/medium/high)
- a short paragraph summarizing the overall direction of changes
- flag any new obligations or liability shifts

This example shows a different document job: comparison rather than extraction. Notice how it explicitly tells the model to ignore cosmetic changes. Without that instruction, the output would be cluttered with formatting differences that obscure the substantive ones.

Prompt block

Summarize this PDF.

Better prompt

I am uploading a policy document.

Task: extract the parts that affect onboarding and access control.
Output format:
1. a 6-bullet summary of the sections that matter most
2. a table of concrete obligations, owners, and deadlines if they exist
3. a short list of ambiguities or missing details I should review in the original document

Do not summarize irrelevant sections just to be complete.

Why this works

The better prompt defines the document's role and the desired extraction pattern, which prevents generic summary behavior. When you tell ChatGPT what the document is and what kind of output you need, you are essentially giving it a lens through which to read the file. Without that lens, the model tries to be comprehensive, which dilutes attention across the entire document. With it, the model can prioritize, skip irrelevant sections, and organize the output in a form that matches your actual workflow.

The request for ambiguities is particularly valuable. It turns the model into a review partner rather than a summarizer. Instead of getting a polished answer that hides uncertainty, you get a structured output that tells you where to look next in the original.