An AI Tutor That Knows Which Course You Are In

The server builds a packet of your real course material and instructs the model to answer from it. Four things go in.

Your class meetings, labs and discussion sections included, since those are stored as their own meeting kinds. Your grading scale, as plain text — A at 93 percent and up, B at 83 and up, or whatever your school actually uses, if you have customized it.

The structured items from your most recent syllabus scan: capped at 8,000 characters and up to 60 items. And your currently tracked deadlines: also capped at 8,000 characters and up to 60 tasks.

Notes are read newest first — the 10 most recent files for that course, sharing a 24,000-character budget of extracted text. A syllabus or deadlines block that has to be cut is explicitly marked as truncated, so the model knows it is working from an abridged source rather than treating a partial list as complete.

• Class meetings, including labs and discussion sections
• Your grading scale, customized cutoffs included
• Syllabus: 8,000 characters, up to 60 items
• Deadlines: 8,000 characters, up to 60 tasks
• Notes: 10 newest files, 24,000 characters shared

This is the constraint that matters most, because it is the one where a confident wrong answer does real damage.

For questions about what is due and when, the tutor answers strictly from your actual tracked tasks. It does not reason its way to a date, and it does not fill a gap with something plausible. If the information is not in what you have given it, it says so plainly and offers general help instead of inventing a specific.

It cites what it used in ordinary language rather than footnotes — "your syllabus lists…", "from your Week 3 notes…" — which is enough to tell you whether an answer came from your material or from general knowledge. That distinction is the whole point.

The deadlines block carries titles, types, due dates, due times, weights, and whether you have checked something off. It does not carry your scores.

So the tutor knows the final is worth 30 percent of your grade and that you have not done it yet. It does not know what you got on the midterm, and it cannot tell you what you need on the final — that is what the forecasting calculators in grade tracking are for.

This is a design decision rather than an oversight, and it is worth knowing so you do not ask it a question it will answer badly. A tutor that had your scores could be more helpful and would also be holding more of your record than a chat feature needs.

Notes attach to a course, so open the tutor from a course rather than on its own — if you have not, the app tells you to instead of accepting an orphan file. PDFs and photos both work.

Extraction happens on the server, not on your phone. The file goes to a private storage bucket filed under your own user ID, and the first time a request needs it the server reads it, extracts the readable text preserving structure, and caches the result so the next question does not pay the cost again.

One limit worth stating: a file over roughly 6 MB is skipped at extraction time rather than sent to the model. If a scanned lecture set is not showing up in answers, size is the first thing to check.

Conversations are scoped to a course, which is what makes the grounding coherent — a thread for organic chemistry is not carrying context from your literature seminar.

The last 12 messages of a conversation are replayed each turn, roughly six exchanges of working memory. That is enough to follow up on an explanation without the thread quietly dragging an hour of unrelated context into every request.

There is also a general thread for questions that are not about a specific course, where it behaves like a competent general assistant with none of the course grounding.

Fifty tutor messages per rolling 24 hours per account — rolling, not a reset at midnight — and 4,000 characters per message, enforced in the composer and checked again on the server.

The model call goes to OpenAI GPT-5.6 Luna with low reasoning and an output ceiling of 2,048 tokens: enough for a worked explanation without inviting an essay. If the provider returns a retryable error the function backs off and retries up to three times, so you see one spinner rather than a failure.

Replies are plain text by instruction — short paragraphs and bullets, no markdown headers — which is a deliberate trade. If what you want is a long formatted document, this is the wrong surface.

The tutor is part of Pro, at $4.99 a month or $29.99 a year, bought with a card on the web or inside the iOS app, and applied account-wide either way.

• 50 messages per rolling 24 hours, per account
• 4,000 characters per message
• 12 previous messages replayed per turn
• Replies capped at 2,048 tokens, plain text

It is for the questions that need your context to answer: what does my syllabus say the exam covers, what did my Week 3 notes call this, what is actually due before Friday, how does my grading scale treat an 88.

It is a poor fit if what you want is a general tutor for a subject you are studying outside a tracked course, because the grounding has nothing to work with. It is also a poor fit if you want long formatted output.

And the same rule applies here as everywhere else in the app: it is worth exactly as much of your semester as you have entered. A course with a scanned syllabus, real deadlines and a few uploaded notes gets useful answers. An empty course gets a general chatbot.