An instructor runs a twelve-week course for a cohort of forty. The reading list is nine PDFs. Lectures are recorded because a third of the cohort is part-time.
Office hours are dominated by three questions that the week-four reading answers on page eleven, and by students who missed a lecture and want to know what was in it.
A study assistant built from the same material replaces none of that. It answers the page-eleven question at eleven at night, which is when students actually have it.
The syllabus is a collection
Individual documents are the wrong unit for a course. A student does not think "which of the nine PDFs". They think "week six".
Upload the readings, then build a collection around them. Tick what belongs, and drag each document into place. That order is precisely what students see in the sidebar, so categories named for weeks or units, ordered deliberately, do your navigational work for you.
Allow chat across all documents decides whether the cohort can query the whole reading list at once. With it on, students can also tick a subset, week five and week six before a seminar, and a small ? beside the menu opens an explainer contrasting the two, which saves you explaining it forty times.
Citations in a multi-document answer name which document each source came from, so a student can always trace a claim back to a specific reading and page.
Recorded lectures become searchable
Audio is a first-class upload type. MP3, WAV, M4A, OGG, FLAC and AAC are transcribed automatically and then trained like any text document.
A lecture assistant is useful in a way a reading assistant is not. It answers "did she say anything about the exam format?" from material that otherwise exists only as ninety minutes nobody is going to re-listen to.
Two practical notes. Voice training requires Business or above. And because a transcript has no page numbers, citations from audio content cannot point at a page that never existed, though they still work.
Long recordings take a while to transcribe, so a ninety-minute lecture is not a thirty-second job. Start it and let it run.
Three settings that change how it feels to study with
Default chat tone. Chatty gives full-length answers with headings and context, which is right for a reading a student is trying to understand. Succinct gives short, scannable replies for quick lookups on a phone. Students switch mid-conversation whatever you choose, by tapping the badge on any answer to redo it in the other style, in place. Their choice beats your custom prompt, which is the correct precedence.
Memory is off by default and worth turning on for a course.
With it on, the assistant uses a signed-in reader's own earlier conversations with that same document as context. A student who explained their situation in week three does not have to explain it again in week nine, and can ask what they discussed last time.
Two boundaries keep it honest. Facts still come only from the document, so memory adds continuity and never new subject matter. And it spans only that reader's own conversations with that one document: never across documents, never across people. Deleting a conversation from their prior chats clears what the assistant remembers from it.
It does nothing for anonymous visitors, which is one good reason to have your cohort signed in.
Key Topics costs nothing and does a lot. It is a ranked list of what a document is actually about, extracted during processing, and every topic is clickable: the question lands in the chat already typed. For a reading nobody has opened yet, it is the fastest way in, and it is on by default. See tuning the chat experience.
Access for a cohort
Course material is usually neither public nor tied to accounts you control, which makes a passcode the pragmatic choice, set at the collection level so one passcode opens the whole syllabus.
It unlocks every document inside, including any whose own level is stricter. The editor warns you before you save, and for a course that is exactly the behavior you want: a student should not meet three unlocked readings and three password prompts. A correct entry is remembered on that device for a long period, so nobody is re-prompted every week.
If you would rather hand access to named groups and be able to take it back, a second cohort or an external examiner, token links give you individually revocable links with optional expiry dates and use limits. See access levels explained.
Students read on phones
On an iPhone or iPad in Safari, the tools list ends with Add to Home Screen, which walks through Apple's own steps and leaves the chat opening like an app.
Two settings on your side make it look right. Your organization's home screen icon, which needs to be a square PNG of at least 512 by 512. And the document's Home screen name, because a home screen shows only about twelve characters. "Ethics W1–12" survives where the full document title does not, and the tool warns you when your name is too long before the student taps Add.
Quizzes, and what the numbers actually say
Generate a question bank from the document. Docutrain sizes it from the content and writes each question with an explanation and page citations. Pro plan and above.
Review it before enabling it. Editing a question creates a new version and retires the old one, so a student's completed attempt still shows the question they were genuinely asked, months later.
From the student's side: pick a length, work through questions drawn at random, and get feedback immediately. Selecting an option locks it, because grading happens on Docutrain's servers and the answer key never reaches the browser.
The review screen is what makes this a study tool rather than a test. A Missed only filter collapses it to what they got wrong, each with an explanation and page chips that show the source passage. Attempts are unlimited and each retake draws a fresh selection. See quizzes.
Afterwards, read two things together.
Quiz statistics give you attempts, average, highest and lowest. The organization-wide Intelligence tab gives you a table of every document with its questions, satisfaction and last activity.
Paired, they tell you something no single number does. A reading with heavy question volume and low satisfaction is hard. A reading with no questions at all was skipped. Those need completely different responses, and without both figures they look identical on a dashboard.
The same logic applies to quiz averages. A low average on material the cohort should know usually points at the source text rather than the students, and it points at a specific section, because the questions carry page citations.
Generating an Intelligence report at the start and the end of a term gives you two comparable snapshots of the same cohort, which is about as close to a controlled measurement as teaching gets.