The single most useful thing you can do with AI while studying is not asking it to explain the topic. It’s pasting your own notes in and telling it to quiz you, one question at a time, refusing to show the answer until you’ve tried. That works offline on a phone with a 4B model, in a library basement or on a train with no signal.
The distinction matters because the two uses pull in opposite directions. An AI explanation feels like learning and mostly isn’t. Being asked a question you can’t quite answer is uncomfortable and is where the learning happens.
Quiz yourself instead of re-reading #
Re-reading a chapter produces a strong feeling of familiarity and weak recall a week later. Retrieving the answer from memory, failing, and then checking is what makes it stick, and spacing those attempts across days beats cramming them into one evening. This is the oldest finding in the study-skills literature and the one students most consistently ignore, because retrieval feels worse while you do it.
A chatbot is a patient quizmaster that never gets bored of re-asking the four cards you keep missing. That’s the whole value proposition. Everything below is a way of getting a model to behave like one.
Five prompts that do the work #
1. The quizmaster. Set this as the chat’s system prompt so it keeps applying as the conversation gets long:
Quiz me on the material below. Rules:
- Ask ONE question, then stop and wait for my answer.
- Never reveal the answer before I attempt it.
- After each attempt: say correct or not, give the correct answer in one
sentence, and note exactly what I missed.
- Mix question types: recall, define, apply, and one "why" question.
- Re-ask anything I get wrong three questions later.
Material:
[paste your notes]2. Flashcards you can export. “Turn the material below into 20 flashcards. One per line, format question | answer. No numbering, no extra text.” That format pastes straight into Anki or Quizlet’s import box, and it keeps the model from wrapping everything in commentary.
3. Two-level explanations. “Explain [concept] twice: first in four sentences for someone who’s never seen it, then as a full-mark exam answer with the terminology.” The gap between the two versions shows you which words you’re using without understanding.
4. Gap hunting. “What does this chapter assume I already know? List the prerequisites, then ask me one diagnostic question about each.” Good for the moment when a subject stops making sense and you can’t tell where it broke.
5. Marking practice. Paste the question, the mark scheme and your own answer, then ask for a mark and the specific lost points. Models are decent at “you never stated the assumption” and unreliable at final grades, so read the reasons and ignore the score.
The rule that keeps it honest: quiz from your material, not its memory #
A local model’s knowledge stops at its training cutoff and it has no way to look anything up. Ask it to recall your syllabus, a specific edition’s page numbers or this year’s exam format and you’ll get fluent invention. Why AI makes things up explains the mechanism.
Paste the notes in and the problem mostly disappears, because the facts are in front of it. Two cases still need checking by hand:
- Arithmetic and formulas. Small models get sums wrong. Check every number against a calculator or your textbook, as in what local AI can and can’t do with math.
- Dates, names and citations. If it produces a reference you didn’t give it, assume it invented the reference.
Working from a PDF or a textbook chapter #
Typing notes out is the slow part. An app that takes a file attachment can extract the text on the device and answer from the matching passages, which turns a chapter into a question bank without a copy of it going anywhere. The mechanics are in how to chat with a PDF offline on your phone.
Long material runs into the context window. A model with a large context window can hold a whole chapter plus your running conversation; a small one will start dropping the earliest part of the chat. If your answers start drifting off the source, split the material into sections and run a separate chat per section.
Which model to use on a phone #
| Model | Size | Good for |
|---|---|---|
| Qwen 3.5 4B | 2.74 GB, needs 3+ GB RAM | Everyday quizzing and flashcards, fast enough to keep a rhythm |
| Ministral 3 3B | 2.15 GB, 256k context | Long chapters and documents in one chat |
| Qwen 3.5 9B | 5.68 GB, needs 6+ GB RAM | Harder explanations and worked problems, slower per answer |
| Gemma 4 E2B | 2.04 GB, needs 3+ GB RAM | Phones with little memory to spare |
Personal LLM runs these open models on the phone’s own chip, so a revision chat containing your half-formed answers stays on the device, with no account and no server behind it. Two features matter for studying specifically: a per-chat system prompt, so the quizmaster rules stay in force for the whole session, and a thinking mode that shows the model’s working before the answer, which is useful for checking a maths or physics step rather than just reading the result. It’s free on iOS and Android.
A practical note on settings: use a lower temperature (around 0.3) when you want strict marking, and a middling one (around 0.7) when you want varied question phrasing. And in airplane mode an ad-supported app can’t load ads, which as a side effect makes offline the least interrupted way to study.
Where offline actually matters #
Study time is mostly dead time: commutes, the twenty minutes before a lecture, a lunch break, a library with concrete walls and no signal. Those are exactly the places a cloud chatbot stalls. They’re also the places where a phone is the only device you have with you.
The other case is a prepaid or capped data plan. Model files are large once, at download, and then nothing. A month of daily quizzing costs no data at all.
Is this cheating? #
Being quizzed on your own notes is studying, in the same category as a friend holding the flashcards. Having a model write the assignment you submit is not, and the line between them sits at what you hand in. Institutions differ on where they draw it for drafting and editing help, and some now require a declaration. Where the line sits on using AI for homework goes through the common policies. If your course has a written policy, that policy wins over any general advice.
Frequently asked questions #
Can AI make flashcards from a PDF without internet? #
Yes, if the app extracts the text on the device and runs the model locally. Attach the file, ask for question | answer lines, then paste them into your flashcard app. Check a sample of the cards, because a model will occasionally produce a confident card about something the document doesn’t say.
Which AI model is best for studying on a phone? #
A 4B model is the practical default: it handles quizzing, summarising and flashcards and answers quickly. Step up to a 9B model if your phone has 6 GB of RAM or more and you’re working through explanations rather than drills. Sizes and memory requirements are covered in how much RAM you need.
Can an offline AI grade my essay? #
It can give you structural feedback, which is the useful part: unsupported claims, missing counter-argument, a conclusion that introduces new material. Treat any number or letter grade as noise. Marks depend on a rubric and a marker, and the model has neither.
Does studying with AI work in a language other than English? #
Small models vary a lot by language. Some are trained on many languages and handle study material in them well, others degrade noticeably. Test with material you can check before relying on it, and see practising a language with offline AI for the language-learning case specifically.
What’s the fastest way to start? #
Take one page of notes, paste it with the quizmaster prompt, and answer ten questions. It takes about five minutes and tells you immediately which parts of that page you actually know.