How to work with Gemini
Gemini is built for work whose input is a file: you attach documents, spreadsheets, images and video, and ask questions about those. It is the only one of the three big chatbots that publishes its exact ceilings as numbers, so you know where you will be stopped before you start.
- Lesson 7 of 15
- Beginner
- Free, no signup
One job with a file, from attaching to a checkable answer
Skip steps two and three and you get an answer that reads well and cannot be checked.
-
1
Attach the real file
A document you actually have and actually need to read.
-
2
Pin the answer to that file
Do not use your general knowledge. Without that line the source of each sentence is unknown.
-
3
Ask for references, not a summary
A section or page number at the end of each line. A summary without references cannot be checked.
-
4
Let it say it does not know
If it is not in the document, write not in the document. A gap is itself information.
-
5
Open two of the references yourself
Citing is not the same as being right. Check two and the rest becomes more trustworthy.
This order is for a job that has one specific file. For a question with no file behind it, this tool has no particular advantage and the order above does not help either.
Last checked: Facts and tool names in this lesson are re-checked against their sources on this date.
The first session: bring a real file, not a question
Someone who tries this tool with a general question gets an answer they would have got from any other tool and concludes there is no difference. The difference starts when you attach something. Start the first session with a document you actually have and actually need to read: a regulation, an invoice, a spreadsheet of three months of sales, a recording of a meeting.
The routine is three steps. You attach the file, you write your question, and in the same message you set the condition that the answer must come only from this file. People drop that third condition and then wonder why the answer contains something the document did not. If you do not say only from this file, the model mixes in its own general knowledge and you have no way of telling which sentence came from where.
One point that pays off in that first session: instead of asking for a summary, ask for references. The sentence "write each point and say at the end which section of the document it came from" produces a summary that can be checked; a summary without references cannot be checked, and something that cannot be checked is no use in professional work.
And before you upload a half-hour video, read section four of this page. The video duration ceiling on the free plan is lower than you think and Google publishes the number, so this is one of the few places where you can plan ahead instead of hitting a wall halfway through the job.
The same durable concepts, plus Gems
Conversation, context and attachment do the same work here that you read about in the ChatGPT lesson: the conversation is the container for one job, context is what is re-read from the top every turn, and an attachment is a file that gets read without becoming permanent knowledge. If you have not read that lesson, read that part and come back; we are not repeating it.
What has its own name here is the Gem: a saved set of instructions that you name and then call up instead of retyping. A Gem is the standing instruction with one important difference, which is that you make several and pick one per kind of job. One Gem for reviewing invoices and one for writing product descriptions, say; each carries its own constraints and tone, and the constraints of one do not land on the other work.
There is a practical decision here that few places mention. An instruction that is right across all your work goes in the general settings, and an instruction that is right for one kind of work goes in a Gem. People who pour everything into the general settings then ask a dry technical question and get the answer back in the voice of a services page.
And one limit that Google states itself and that is useful for planning: the space taken by activity in these apps has a ceiling, and that ceiling is separate from your Google Drive storage. So you can have room in Drive and still get an out-of-space error here, and the fix is deleting old activity in this app rather than emptying Drive.
Four places, and the rule for choosing between them
-
General settings
What is true across all your work: who you are and what it must not say.
-
A Gem
The instruction for one kind of job. You make several and pick one each time.
-
The conversation
The constraint of this one job. Re-read every turn and with you to the end of the chat.
-
The attachment
Data, not instruction. Up to 10 files per prompt, and gone in the next conversation.
The rule in one sentence: whatever is true across all your work moves up, whatever is true for one job stays down. Putting an instruction in the wrong place shows its damage on a job that had nothing to do with it.
Five jobs built on an attached file
All five start with a file, because without one this tool does what the others do. Swap what is inside the braces for your own.
One. Turn a supplier invoice into a usable table. Attach the photo or the PDF. The last line of the prompt is the most important one.
The attached image is an invoice. Extract its rows into a table with the
columns: item description, quantity, unit price, total price.
For any number that is not clear in the image, write "not legible" rather
than guessing.
Recalculate the grand total yourself, and if it does not match the total
written on the invoice, state that difference separately.Two. Ask three questions of a three-month sales file. Attach the spreadsheet. Note that you write the questions; asking for analysis without questions produces marketing copy.
The attached file is my sales for the past {three} months. Answer these
three questions and for each answer say which column and which rows you
computed it from:
1) The best-selling and worst-selling item in this period.
2) The month with the largest change against the month before it.
3) Items whose sales were zero in all three months.
Make no forecast and give no recommendation. Only what is in the file.Three. Turn a meeting recording into decisions and tasks. Check the audio duration ceiling in the next section before uploading; if the meeting is longer, attach the transcript rather than the audio file itself.
The attached file is a recording of a work meeting. Give three lists:
1) Decisions that were made.
2) Tasks assigned to someone, with that person name if it is said in the file.
3) Things that were raised and did not reach a decision.
Where the audio is unclear write unclear, and do not guess that part.
Do not write a date or an amount that is not said in the file.Four. Write a description from photographs of the product itself. For a seller who has no written spec sheet and only has photos.
These {four} images are of one product. First list what you can see in the
images that a buyer would care about.
Then write a {70} word product description that refers only to those visible
things.
Write nothing about material, durability, country of manufacture or warranty;
none of that is visible in a photograph.Five. Get one specific question answered from a long document, with references. This is the job that takes the most time and returns the most.
Read the attached document and answer only this question: {your question}
Answer shape:
- One point per line, with the section or page number in brackets at the end.
- If the answer is not in the document, write "not in the document" and stop
there.
- If the document says two different things in two places, give both and say
where.
Do not use your general knowledge. Only this document.Do not use your general knowledge is the one line all five need. Without it, the answer is a blend of your document and what the model read beforehand, and you have no way of separating the two. With it, every sentence of the answer either carries a reference or declares that it was not in the document, and that single change turns the output from readable text into something you can rely on.
Google has written the ceilings down, so know them in advance
This section is the main reason this lesson exists. Google own help page on uploading files publishes the ceilings as numbers, and everything you read below was read from that page on 6 September 2026. None of these figures is a guess of ours.
Up to 10 files can be added to one prompt, and if you upload a folder, every file inside it counts separately. Each video up to 2 GB and the other supported types up to 100 MB. One code folder or one GitHub repository, up to 5,000 files and a maximum of 100 MB. A ZIP file up to 10 files and up to 100 MB, with video and audio not accepted inside it.
And now the two numbers that matter most and surprise most. Total video length in the base state is up to 5 minutes, and upgrading to Google paid plans raises it to one hour, that is 60 minutes. Total audio length is up to 10 minutes, and with the upgrade up to three hours, that is 180 minutes. A forty-minute meeting you recorded will not upload at all without a paid plan.
Separately, Google writes that there are rolling limits on uploading and analysing files: after a while the ceiling resets and you can carry on. It also gives a number for the context window: the paid plans have a one million token window, which Google itself explains as up to 1,500 pages of text or 30,000 lines of code, and the Deep Think mode has a 192 thousand token window.
| Item | Base state | With a paid Google plan |
|---|---|---|
| Files per prompt | Up to 10 | Up to 10 |
| Size of each video | Up to 2 GB | Up to 2 GB |
| Size of other files | Up to 100 MB | Up to 100 MB |
| Total video length | 5 minutes | 60 minutes |
| Total audio length | 10 minutes | 180 minutes |
The practical consequence of this table is one sentence: for anything longer than these ceilings, do not upload the media file, upload its text. Get a one-hour meeting transcribed somewhere else and attach the transcript; that text is a file of a few hundred kilobytes and none of the ceilings above comes near it. It is the simplest thing you will learn in this lesson and it saves the most time.
Privacy, human reviewers, and where Iran stands
Google writes this part with a directness you do not often see, and it is better to read its own sentence: in the privacy notice for these apps it says that human reviewers, including trained reviewers from its service providers, review some of the data collected, and it asks users not to enter confidential information they would not want a reviewer to see or Google to use to improve its services.
Three settings and one number you should know. The main setting is in the apps activity section, and that is also where you review and delete your activity, change the auto-delete period, and control whether your data is used to improve Google AI. The auto-delete period is 18 months by default and can be set to 3 months, 36 months or indefinite. Someone who has never opened that setting has today conversations sitting there for another year and a half.
And the detail that matters and gets said less: conversations that a human reviewer has seen are not deleted when you delete your activity. By Google own account they are retained, along with their related data, for up to three years. So the delete button does not delete everything, and someone who thinks it does is deciding on a false premise. Google also writes that even when activity saving is off, it still uses chats to respond to you and to protect the service, including with help from human reviewers.
On Iran there is an apparent contradiction worth knowing. Google own page says the Gemini web app is available in more than 70 languages and over 230 countries and territories, and Farsi is on the supported-languages list. But Iran is not on the country list. The language is supported and the country is not, and those are two separate things. We wrote this up with its source on the Gemini page in our AI reference, and the payment layer separately in the guide to buying AI tools. We do not sell accounts and we do not suggest ways around anything.
Four things to know about your data
-
Auto-delete, 18 months by default
It can be set to 3 months, 36 months or indefinite. Untouched, it stays at 18 months.
-
Human reviewers
They review some of the data, including reviewers from Google service providers.
-
Deleting does not delete everything
A reviewed conversation is retained for up to three years, even after you delete your activity.
-
Language yes, country no
Farsi is on the supported-languages list and Iran is not on the country list.
None of these four replaces the main rule, and Google writes the same: do not enter confidential information you would not want a reviewer to see.
The fast path, with AI
If twenty invoices or ten contracts or four same-shaped reports land on you every month, writing a prompt for each one is the same manual work wearing a new face. What professionals do is build one Gem once, holding the extraction schema and the prohibitions, then feed the documents to it in batches, and finish with an audit question about which fields came back empty. That last question is what turns a pile of extractions into a work list.
- Create a Gem and put the recipe below inside it. Name it after the job it does rather than the tool, say "supplier invoice extraction", because six months from now you will have four Gems and will have to find them by name.
- Feed the documents in batches, at most 10 files per prompt, and open a separate conversation for each batch. Smaller batches are better: when an error turns up, you need to know which batch and which file it was in.
- For anything long and media-shaped, do not hand over the audio or video file, hand over the transcript. See the ceilings in section four: a total of 5 minutes of video and 10 minutes of audio in the base state, and a transcript meets neither of those ceilings.
- At the end of each batch, ask the audit question: which fields stayed empty for which document, and why. The output of that question is exactly the list of documents you have to look at by hand, and it is usually far shorter than the batch.
Copy-ready recipe
Your role is extractor, not analyst. For every file you are given, fill in these fields:
{document number} | {date} | {counterparty} | {total amount} | {due date} | {subject}
Fixed rules:
- Only from the file itself. Do not use your general knowledge.
- Any field not in the file: write "not in the document". Guessing is forbidden.
- Any field present but unreadable: write "not legible".
- Write numbers exactly as the document writes them and do not convert units.
- If one document carries two different figures in two places, give both and say where.
- Write no summary, no recommendation and no opinion.
Output: one table, one row per file.
After the table, also give this:
Audit: for each document, which fields stayed empty and whether the cause was
missing information or illegibility. If a document was not of this type at all,
say so there.Before you trust the output: Three places where this goes wrong. First, the method makes extraction fast and guarantees nothing about accuracy: compare two rows from every batch against the document yourself, and do not drop that step even after ten batches in a row came back correct. Second, the Gem schema is fitted to the shape your documents have today; the day a supplier changes its invoice layout, the Gem quietly fills the fields with "not in the document" and you will not notice until you ask the audit question. Third and more important than either, these documents are not yours: invoices and contracts carry other people names, numbers and amounts, and higher up this page you read that human reviewers see some of the data. Before the first batch, decide whether you are doing this with real names or with placeholders.
AI in this kind of work
The subject of this lesson is itself an AI tool, so the right question is which one to reach for on which job. Our position: wherever the job starts with a file, this one has a real advantage because you know its ceilings in advance; wherever the job starts with a long text edited many times over, somewhere else runs more smoothly. The pages below are where we keep the access and pricing facts with their sources.
Tools that actually help
- Gemini The subject of this lesson. The only one of these three that publishes file-count, size, duration and context-window ceilings as numbers. Farsi is on Google supported-languages list and Iran is not on the country list; both on the same help page.
- Claude When the output of the job is a long text you edit several times, the work runs more smoothly there. Unlike Google, Anthropic publishes no number for its free-plan usage ceiling; what it does publish is written up with sources in the matching lesson.
- ChatGPT The most common choice and enough for short, frequent jobs. We make no claim about the free ceiling or consumer-app access from Iran, because no OpenAI consumer page opens from this server; the link goes to the vendor page and the reason is written there.
Where it backfires
The specific risk in this lesson is that the documents you attach are usually not yours. A supplier invoice, a client contract and a meeting recording carry other people names, numbers and amounts, and you have no permission to publish them. In the privacy notice for these apps Google itself writes that human reviewers see some of the data collected and asks users not to enter confidential information they would not want a reviewer to see; the same notice says a reviewed conversation is retained for up to three years even after you delete your activity. Put together, that means the decision is made before the upload and not after it. The practical route is simple: replace names and numbers with placeholders, unless the job is meaningless without them.
Sources: Google: Gemini Apps privacy notice Google: upload and analyze files in Gemini Apps Google: where you can use the Gemini web app
Where this advice stops
Every number in this lesson comes from two Google help pages, and those pages change whenever they like; the verification date is printed at the top and after that date this text is a photograph. Google itself notes beside the file-count ceiling that its availability is not guaranteed, so these figures are ceilings and not promises. Second, what is taught here is concepts and routines, not where the buttons are; this product interface has changed several times in a year and will again. Third, this lesson is about work that starts with a file; for a question with no document behind it, none of these five routines gives you any advantage. And fourth, if you cannot reach this service from Iran, none of it is any use to you, and the right move is to spend your time on a tool you can actually sign into.
From our own work
One thing we know from this site own configuration, and you can see it yourself. The robots.txt of this domain has a group listing 16 AI crawlers by name, and one of them is Google-Extended, the crawler that governs Google use of a site content for its AI. The lesson we paid for is this: in robots.txt a named group replaces the star group entirely and inherits none of its rules. As long as that group carried only one allow line, the named crawlers had no restrictions at all, and they were exactly the ones crawling this site infinite login and signup trap; our own note in the code records 8 August 2026, 10 GB of log and a full disk. Now every rule is repeated verbatim inside the AI group too. The point that stays with you: allowing or blocking Google-Extended is the site owner decision and it is made in that file, not in the Gemini account settings.
Real follow-up questions
Why will my half-hour video not upload?
Because the ceiling on total video length in the base state is 5 minutes and reaches 60 minutes with a paid Google plan. File size is not the problem: each video is accepted up to 2 GB and the issue is duration, not size. The practical route is uploading the transcript instead of the video itself.
How is a Gem different from the general settings?
The general settings go into every conversation, and you call up a Gem yourself for one kind of job. The rule for choosing is simple: what is true across all your work goes in the general settings, and what is true for one kind of work goes in a Gem. Pour everything into the general settings and even a dry technical question comes back in the voice of a services page.
If I delete my activity, is everything deleted?
No. By Google own account, conversations that a human reviewer has seen are not deleted along with your activity and are retained, with their related data, for up to three years. The auto-delete period for the rest of your activity is 18 months by default and can be set to 3 months, 36 months or indefinite.