Files and documents: upload and analyse

In the ten-jobs lesson, contract text reached the chat by hand: copy, paste, hope nothing got lost on the way. There's a nicer route — upload the whole file. A forty-page PDF, a spreadsheet, a scan, a slide deck: one move, and the AI sees the entire document, not just the piece that fitted on the clipboard.

What it looks like

There's a paperclip in the input box (or a "+"). Tap it, pick a file off your computer or phone, and write what you want done with it. Not "here's a file" into the void, but file + a specific job: "here's my rental agreement — list my obligations and the penalties"; "here's the sales table — which month was best and by how much".

What you can upload

The list is wider than most people think — which is exactly why it's worth reading to the end.

• PDFs and Word documents — contracts, manuals, course notes, articles.
• Spreadsheets (Excel, CSV) — exports, reports, lists. The AI will total things up, find the maximum, spot the outlier.
• Images and scans — a photo of a page, a screenshot of a conversation, a snap of a bill (that's the bridge from the images lesson).
• Presentations — ask for a summary, or for the weak slides.

Three scenarios where this saves hours

Not abstract benefit but three concrete jobs that used to take an evening.

• A contract. Thirty pages land in your inbox before signing. Upload → "explain in plain words: what am I obliged to do, what's my risk, are there hidden charges, what are the deadlines". A few minutes instead of an evening with a legal dictionary.
• A manual. An 80-page guide to a gadget or a piece of software. Upload → "how do I set up X? answer only from this document". The answer from the right page instead of a long flick through.
• A spreadsheet. A year of expenses exported. Upload → "which three categories took the most, and what percentage of the total". Analysis done, no formulas written.

The "answer only from the document" move

This is the most valuable sentence in the lesson. Out of habit, a language model tops its answers up with whatever it "generally knows". When you need the answer strictly from your file, say so: "use only the uploaded document; if the answer isn't in there, tell me". It cuts the invention dramatically and turns the AI from an imaginative know-it-all into a careful research assistant. Remember hallucinations (plausible inventions) from the 10-jobs lesson? This is how you rein them in.

Hold on — is this safe?

Before you read on, think back to the privacy lesson and answer for yourself: is every document worth uploading?

No. The file goes to the service's servers, which means the same rule as for text: don't upload other people's personal data, trade secrets, or scans of passports and cards. The Samsung story from the settings lesson was exactly this — engineers who "just let the AI help" with internal code. Your own rental agreement, your course notes, your expense sheet — upload away. For secrets: either anonymise them (swap names and numbers for "X") or don't upload at all.

Limits: what not to expect

The honest part: here's what files and documents can't do, however you ask.

• Very large files may not be read in full — for a 500-page book, upload it chapter by chapter and ask in pieces.
• A bad scan (crooked, blurry) reads worse — shoot it square-on in good light.
• Double-check the summary of anything important yourself: a contract clause with money riding on it deserves your own eyes. The AI speeds you up; the responsibility stays yours.

When it isn't one job but two hundred

Everything above is about one document. Plenty of people have a different problem: not writing an email but producing two hundred product descriptions, two hundred replies to reviews, two hundred captions. One at a time in a chat isn't work, it's a punishment.

In batches it goes like this.

1. Get the template right on one item first. Make one description exactly what you want, and only then multiply. Otherwise you multiply your own mistake two hundred times.
2. Fix the template in the project instructions so you don't repeat it in every message. Projects were the previous lesson.
3. Upload the spreadsheet with your source data: name, specs, price. Ask for ten to fifteen rows at a time, not all of it at once.
4. Spot-check every batch. After a few dozen items the model starts drifting from the structure: lengthening here, reordering there, adding something the data never had. Only your eyes will catch that.
5. Take the result back as a table — ask for rows with separators so you can paste them back rather than retyping.

One thing about invention: in a batch it's more dangerous than one at a time. In a single text you'd notice a made-up spec; in the two-hundredth you won't. So put it in the project instructions directly: «use only what's in the table; if the data is missing, leave the field empty and invent nothing».

The bridge back to tidy chats

Remember projects from the order lesson? A file you'll work with more than once (course notes, a client's brand book) belongs in a project rather than a one-off chat — then it's at hand in every related conversation and you never upload it twice.

How to ask a good question about a document

Uploading the file is half the job; the other half is asking properly. A weak "tell me about this document" gets you a watery summary. A strong one sets the role, the job and the format. Compare: "read this contract as a lawyer on my side: put my obligations, penalties and deadlines in a table, and at the end flag the three clauses where I'm most exposed". Same file, but the answer is now a working tool.

The formula from the first-dialogue lesson is back: situation + constraints + the result you want. And if the document is long, don't hesitate to fire a series of questions at it in the same chat — the model holds on to the uploaded file for the whole conversation, the way it held on to that sofa.

Do this now

Find a real document on your device — a contract, a manual, a bill, a spreadsheet — upload it and ask one specific question with "answer only from this file" tacked on. Then weigh up how much time that saved against reading it yourself.