Set up AI for the new school year

It is the first week. Your downloads folder holds four course outlines named things like syllabus_FINAL_v3.pdf, there is a chapter to read by Thursday, and two teachers have said opposite things about AI: one banned it, one wants you to use it and declare it. You need to know which uses are allowed in each course.

Ten minutes of setup now saves you a term of pasting things into a blank chat and hoping.

1. Make one folder per subject

Create a folder for each course on your timetable. Chats you start inside one stay there, so in week nine you can find the thread where mitosis finally clicked instead of scrolling past two hundred conversations.

The free plan allows 25 folders, more courses than anyone is taking. Where the button lives depends on the device: in the mobile app the plus beside the Folders heading works on any signed in account, while on the web the New Folder button currently appears on Pro accounts. On a free account, make the first folder in the app and it will be there when you open the web version.

A folder sorts your chats and nothing more. A chat inside "Chemistry" will not produce more chemical answers.

2. Fill in Personal preferences once

Open Settings and find the Personal preferences box, called Personalization in the mobile app. What you put there is passed into every new conversation. Four lines is plenty:

Your level. "Second year undergraduate, first time doing statistics."
Your courses and the textbook each one uses.
How you want to be taught. "Explain the idea, show one worked example, then give me one to try."
The rule you want enforced. "Never write text for me to hand in. Ask me questions instead."

That last line does more work than people expect. A model told to quiz you will quiz you.

Free plan note. Saved memories, the ones that build up about you as you chat, are a Pro feature at $20.99 a month. Personal preferences works differently: you write it yourself, once, and it is on the free plan. For a student who is not paying, it is the whole setup.

3. Match the model to the task

MultiChats carries 60 models from 18 providers. The right one changes by subject. On the web the search and reasoning buttons named below appear once you switch Intent Detection off in Settings, under Tool Preferences; on by default, it reads each message and decides for you. The mobile apps always show them.

Maths, physics, proofs. Use a model that reasons before answering: GPT-5.6 Sol, Claude Opus 5 or Gemini 3.7 Flash. On the free plan, GPT-5.4 Nano and Gemini 3.5 Flash Lite reason. Pro adds a reasoning effort control, Off up to High.
Essays, translation, close reading. Claude Sonnet 5 and GPT-5.5 handle structure and tone well, both on Pro. Ask them to critique your draft, never to supply one.
Handouts, past papers, scanned chapters. Claude, Gemini, Grok 4.3 and the current GPT models read PDFs, most open weight models do not, and the picker badges the ones that can.
Checking a fact or a date. Web search runs on the Claude and Gemini models. On the free plan that narrows to the Gemini Flash Lite entries, with 2 searches per 30 days, and in the mobile app the search button is Pro only.

You can switch models mid-conversation without losing the thread, so when step three of a proof still feels like magic, ask the next model to explain it from scratch.

4. Try it on one real document first

Do not wait for a graded assignment to see how this behaves. Upload one course outline, ask for every date and deliverable as a list, then check that list against the PDF line by line. Week one is a better time to find out what a model gets wrong than week eleven. Free accounts get one file per message, up to 10MB, three uploads a day.

Watch the length. The free models that read PDFs top out around 100,000 tokens of working context, roughly 75,000 words, and some sit lower. If answers start ignoring the first half of a document, split the chapter across two chats. Here is what a context window is.

5. Save the three prompts you will reuse

Keep these in a phone note.

Explain, then test me. "Explain [concept] at second year level, then ask me three questions about it, one at a time, waiting for my answer each time."
Mark my draft against the rubric. Paste the rubric, then your draft. "Grade this and quote the sentence that costs me marks under each criterion. Do not rewrite anything."
Find the weak point. "Here is my argument in three sentences. Which would a marker attack first, and why?"

All three ask the model to do something to work you have already done. That distinction is what the next section turns on.

Where the line is, for students and parents both

Most schools and universities draw the line between understanding and production. Using AI to understand something is treated much like a tutor or a study group. Using it to produce what you hand in counts as someone else doing your work.

On the safe side:

Having a concept explained again, differently, until it lands.
Being quizzed on material you have already read.
Feedback on a draft you wrote, in the marker's terms.
Turning your own messy notes into a revision sheet.

Not on the safe side:

Generating prose, code or working that you submit as yours.
Asking for a bibliography and pasting it in, which is also a factual disaster for the reason below.
Handing in an answer you could not defend if a tutor asked you on the spot.
Using it where the instructions say no AI.

One test covers most cases. Close the app and explain the thing to someone. If you cannot, you have not learned it, and what you are about to submit is not yours either.

Two habits. Read your institution's policy this week, because departments inside one university differ. And keep a one line record of what you used AI for on each piece of work, in case your course asks for a declaration.

For parents: the conversation worth having is about the line above. Agree where your household draws it, write that agreement into Personal preferences so the rule travels into every chat, and ask to see the one line record on each assignment.

What it is genuinely bad at

Citations. With search off, a model works from memory, and memory invents plausible references: a real author, a real journal, a paper nobody wrote. Never cite something you have not opened. With search on, answers carry citation pills that link to the source, and clicking through is the point. Long arithmetic. Reasoning models are far better than they were, but a multi step calculation can still carry a unit error through to a confident final number. Check the last line yourself.

Your marker. It does not know your department's house style or how your rubric is weighted. It guesses at both, fluently.

Anything recent. Without search, a model's knowledge stops at its training cutoff, so check this year's exam format yourself.

Before the first deadline

Do this while nothing is due. The version of you at 11pm the night before a hand in will not stop to make folders, and that is the night the rules matter most.

MultiChats is free to start: 50 messages a day, 21 models. Pro is $20.99 a month for the full picker, saved memories and the reasoning effort control. See what one subscription across models covers.

Is it cheating to use AI for homework?

Being quizzed, having a concept explained, and getting feedback on your own draft are usually treated like tutoring. Generating the text you submit is not. Read your institution's policy.