DESK · THEORY
WorkflowBeginner · June 4, 2026 · 9 min read
On this page

Paste a screenshot, upload a PDF: working with documents and images in AI

Stop retyping information that already exists. Three inputs, plain chat, no setup.

What you'll have when you're done

A working pattern for feeding Claude documents and images directly, so you stop transcribing what is already written down. A contract summary you did not have to read line by line. A screenshot of a chart explained in plain language. A spreadsheet export turned into a plain-English paragraph about what is actually in it.

No terminal. No API. Plain chat, in the browser.

The CEO who retyped the table

I used to retype things all the time. Pulling numbers from a vendor PDF into a slide. Squinting at a screenshot of a dashboard and typing its contents into a chat prompt so I could ask a question about them.

You do not have to do any of that. Claude reads documents and images directly. You give it the file or paste the image; it reads it and responds as if the content were already in your prompt. This is one of the most underused moves in plain-chat Claude.

What you need first

That is it. No technical setup.

Step by step

Three inputs. Each works the same basic way: give Claude the file or image, tell it what you need, and go. The differences are in what each input type is good at and where it trips up.

Paste a screenshot or photo

Best for: pulling text out of a screenshot, explaining what a chart or dashboard shows in plain language, turning a whiteboard photo into structured notes.

You have a photo of a whiteboard from your team offsite. You have a screenshot of a competitor's pricing page. You have a slide from a vendor deck that was emailed as a PNG. You have a picture of a handwritten page someone faxed you (yes, faxes still happen).

Paste it directly into the chat input, add a prompt, and send. That is the whole move.

Good prompts for screenshots:

Where it gets unreliable: low-resolution or small-text images. If the image is blurry or the text is tiny, Claude will do its best but the transcription will have errors. Dense charts with lots of data points are also risky for exact numbers. It can misread a value off a busy graph. The principle here is the same one that applies everywhere in AI work: Claude sounds confident whether it is right or not. Never ship a number it read off an image without going back to the source and checking. Use it for structure and plain-language explanation; verify any load-bearing figure yourself.

Upload a PDF

Best for: summarizing long documents, answering targeted questions about what a document says, and pulling out key terms, risks, or dates.

Open a conversation, click the attachment icon, and upload the PDF. Then ask your question directly.

Good prompts for PDFs:

A contract review that used to mean three hours and three sticky notes can come back in ninety seconds as a bulleted summary with the flagged language quoted. You still want your lawyer to look at anything with teeth. But arriving at that call already knowing the document saves everyone time.

Where it gets unreliable: scanned PDFs (photo-quality, no text layer). If a PDF was printed and re-scanned, Claude is reading a photograph of text, and the same cautions from the screenshot section apply. Dense financial tables in scanned documents are especially error-prone.

Very long PDFs are also a real limitation. For a document that runs to many hundreds of pages, Claude may skim or miss details in sections it skipped. If a specific clause matters, ask about that clause by name rather than trusting a general summary to surface it.

Verify every load-bearing figure against the original before it drives a real decision. "What does the revenue-share provision say?" is a good question. "Add up all the financial obligations in this contract" is a question where you check its arithmetic, not trust it.

Before you upload: check your data settings and your company's policy first. Contracts, employee records, and confidential financials are real-world documents with real-world consequences. The free tier has less conservative data handling than the paid plans, and your company may have its own rules. Handle sensitive documents accordingly.

Drag in a spreadsheet or CSV

Best for: a plain-English summary of what is in the file, spotting structural issues (missing values, duplicate column names, data that doesn't look right), and drafting a narrative around numbers you already understand.

Export a CSV from your CRM, your accounting tool, your analytics platform, or wherever the data lives. Drag it into the chat window or use the attachment button. Then ask what you need.

Good prompts for spreadsheets and CSVs:

Where it gets unreliable, and this one matters: plain-chat Claude reads a spreadsheet as text, not as a live spreadsheet. It does not calculate. It reads the numbers the way you would read a printed page.

Do not ask it to sum revenue across rows, calculate a growth rate, or derive a total. It can produce a number that looks plausible and is wrong. Treat any figure Claude computed from a file as a draft to verify, not a result to use.

Multi-tab files are also risky. If your workbook has five tabs and only the first is read, Claude will not know what it missed. Ask "what sections did you see in this file?" to find out what it actually ingested.

If your goal is to do math on the file, the Claude Excel add-in writes real formulas in real cells (more on that under Level up). Plain-chat spreadsheet work is for orientation and narrative, not calculation.

How you'll know it's working

Claude's response used the actual content of your file. Not generic observations about documents in general, but specific details from your specific PDF: the name of the parties, the actual dates, the quoted language. If the output sounds generic, the file probably did not load correctly. Try re-attaching.

You caught something you would have missed. A clause you skimmed past. A column in a spreadsheet you forgot was there. A number on the chart you would have misread in a hurry. That is the real value: it reads more carefully than you do when you are under time pressure.

You used the output. Forwarded the contract summary to your lawyer with the three flagged clauses noted. Pasted the chart explanation into the board update. If you looked at it and closed the tab, the task was not real enough.

When it breaks

Claude says it cannot read the file or returns nothing useful. Scanned PDFs with no text layer are the most common culprit. Try a different export if you have one. For a photo-heavy report, pasting key pages as screenshots sometimes works better than uploading the whole PDF.

The output has wrong numbers. Expected, and serious. Do not trust any figure Claude pulled from a document or image without checking the source. Build the habit: anything computed or read off an image or table gets verified before it travels.

It only read part of the document. Ask: "What sections of this document did you have access to?" If it missed something critical, pull that section into a separate file and feed it in directly.

The image was too low-resolution. Get a higher-resolution version if you can, or type the relevant text manually.

You are not sure what Claude has access to. Claude only knows what you gave it in this conversation. A referenced appendix you did not upload does not exist for it. A table cut off at the edge of your screenshot does not exist for it. Ask: "Is there anything in what I gave you that you could not read?" and it will tell you.

On privacy: Claude does not know about a document unless you share it. But once you paste something in, it is in the conversation. For contracts, personnel files, or anything confidential, check what is safe to share with AI tools before you upload.

Level up

The three moves above work in plain chat. No setup, no integration. When you are ready to go further:

If you are working with financial models, Claude inside Excel is the right next step. The plain-chat move gives you orientation. The plugin lets Claude write formulas and restructure tabs directly. The jump in capability is significant.

If you want to ask questions across a P&L over time, ask your P&L anything covers a setup where your financial data is always available, not just on the days you remember to upload it.

If you want documents available across every conversation, Claude Projects lets you upload context once and have it present in every session inside that project. Add your standard agreement once. Every vendor review conversation inherits it.

Start with the move that fits the pile on your desk right now. Upload one PDF you have been avoiding. See what comes back.

The Thursday 3

Get three workflows like this every Thursday

The Thursday 3 is a free weekly email. Three workflows that put you in the top 1% of CEOs. 90-second read. Every card links back to a step-by-step guide like this one.

Get the newsletter →
The Desk Theory books

The architecture behind this workflow.

Two operator manuals for the same job, run two ways: OpenCLAW for the always-on harness, Claude Code for the focused-work CLI. Pick one, or get the bundle for $149.

Browse the books · $99 each

Want one workflow like this taken apart end-to-end every week? The Tuesday Pro Deep Dive · $39/mo.