Running a course on Uedu — Uedu handbook

1. Before you start

What you need

A browser. Nothing is installed, by you or by your students. Recent Chrome, Edge, Firefox or Safari.
An institutional email address. You can sign in with anything, but claiming or creating a course requires verification against an academic domain (.edu, .edu.tw, .ac.jp and similar). Personal addresses such as Gmail are not accepted for that step.
An authenticator app on your phone. Two-factor authentication is mandatory for teacher accounts — see below.
About two hours for a first course set-up: roughly one hour writing the assistant's instructions and one hour uploading material.

Creating an account

Go to uedu.tw/signin and sign in with Google, Apple, GitHub, a passkey, or email and password. If your institution has its own entry point (yourschool.uedu.tw) use that instead — it is the same platform with your institution's branding, and every account and course works identically either way.

Two-factor authentication is not optional for teachers

Every account with teacher, teaching-assistant or administrator rights must enable 2FA, and other features stay restricted until it is set up. The reason is blunt: a teacher account can read student dialogue. Set it up during the same sitting in which your course claim is approved, and store the recovery codes somewhere that is not the same phone.

Interface language and time zones

The interface is available in 16 languages; the switcher is in the header and the setting follows your account. Students choose their own independently of you.
Students converse with the assistant in whatever language they write in, regardless of the interface language.
Times are stored in UTC and rendered in each viewer's local time zone. A deadline you set at 23:59 your time shows correctly to a student in another country — you do not need to calculate an offset, and you should not try to.

2. Getting a course

A course is the unit everything hangs off: the assistant, the material, the coursework, the analytics. There are two ways to get one, and both end in a short human review that confirms you are the teacher you say you are.

Path A — claim an existing course

If your institution's catalogue has been imported, your course is probably already in the system and you claim it rather than create it.

From the sidebar: Teaching toolkit → Course setup → Claim a course.
Search by institution, term and course name or code, then claim the matching entry.
Verify your identity: enter your institutional email, receive a six-digit code, enter it. The code is valid for 15 minutes, and there is a five-minute cooldown between sends.
Submit: your name, your unit, and optionally a document that supports the claim (teaching certificate, appointment letter, a screenshot of your timetable — PDF, DOC, DOCX, JPG or PNG, up to 5 MB). Optional, but it speeds up review.
Wait for the approval email. You can check progress under My claim requests.

Path B — create a new course

If the catalogue does not have your course — which is the normal case for an institution that has not sent us one — create it. Same verification step, same review.

Required: course name, institution, term. Everything else can be edited later.
Paste your syllabus into the description field. If you do, the system drafts an initial set of assistant instructions from it, which is a much better starting point than an empty box. This is the highest-value optional field on the form.

Once approved

The course is created private by default, with a six-character join code. Nothing is visible to anyone until you hand that code out. The approval email contains the student join link.

3. Writing the assistant

The assistant's instructions — its system prompt — are the single setting that determines whether students find it useful. It is free text in any language, it is not code, and it is editable at any time during the semester.

Where

Teaching toolkit → Teacher console → your course → Teacher dashboard → Course management → Course settings.

What to put in it

Four elements, in this order, cover most of what matters:

Role. "You are the teaching assistant for Introduction to Political Theory, a first-year undergraduate course."
Course facts. Level, main texts, what has been covered so far, anything the assistant would otherwise guess wrong.
How to answer. Depth, worked examples, when to ask a clarifying question, what language to reply in if you want one fixed.
Boundaries. What it should refuse or redirect — "do not write the essay for the student", "for administrative questions about deadlines, tell them to ask me".

You are the teaching assistant for "Introduction to Political Theory", a first-year undergraduate course taught in English. Course texts: Plato's Republic (Books I–IV), Hobbes' Leviathan (Chapters 13–18), Rawls' A Theory of Justice (Chapter 1). How to answer: - Explain in plain language first, then in the vocabulary of the text. - When a student states a position, ask what the strongest objection to it would be before agreeing. - Quote the text when you rely on it, with the book and chapter. Boundaries: - Do not draft or rewrite a student's essay. Comment on structure and argument instead. - If asked about grading, deadlines or my policies, say you do not know and tell the student to ask me directly. - If a question is outside the course, answer briefly and say it is outside our scope.

The default is "answer well", not "be Socratic"

Course assistants answer directly by default. That is deliberate: a student stuck at 1 a.m. needs an answer, and constant indirect questioning from a course assistant reads as evasion. Socratic questioning is available as an explicit mode (chapter 6), and the independent learning companion, Aida, is Socratic by design. Choose the behaviour you want rather than inheriting one.

Common mistakes

Too long. Two thousand words of instructions produce worse behaviour than three hundred well-chosen ones. Write less, then add rules when you see something you dislike in the actual dialogue.
Contradictory rules. "Never give answers" plus "be helpful and complete" resolves unpredictably. Decide which wins and say so.
Facts that belong in the material. Do not paste your reading list contents into the prompt. Upload the readings (chapter 4) — then it can cite them.
Writing it once. Read a sample of real dialogue in week two and revise. It is the fastest improvement available and almost nobody does it.

4. Adding your own material

Uploaded material is what turns a general model into a course assistant. When a student's question relates to something you uploaded, the relevant passages are retrieved and used, and the reply cites the file and page so the student can check it against the source.

How to do it

Teacher dashboard → AI knowledge base → upload. PDF, DOCX and PPTX are the normal formats.
Processing runs in the background; a large file is usable within minutes, not instantly.
Upload the readings students actually work with — lecture slides, chapters you have rights to distribute, your own notes, past worked examples.

Two things that silently do not work

Scanned PDFs with no text layer. A photographed book chapter is an image as far as retrieval is concerned. If you can select text in your PDF viewer, it will work; if you cannot, run OCR first.
Material you may not distribute. Uploading is distribution to your class. Copyright is your responsibility, exactly as it is on your institution's own LMS.

What it enables downstream

Grounded, cited answers in dialogue.
Quiz and worksheet drafting from your actual content rather than from the model's general knowledge.
The course knowledge graph, which extracts concepts and their relations from the material and then tracks which concepts each student has demonstrated understanding of. It cannot be generated without uploaded material.

Exact retrieval parameters — chunk size, embedding model, top-k — are documented for methods sections in the retrieval methodology note (Chinese).

5. Choosing a model

Most courses never touch this page and are fine. Read this chapter only if you have a specific reason to.

Option	Needs an API key?	When to use it
Platform default (free tier)	No	Everything, unless you have a reason otherwise. Enabled automatically in every course; it is what students get if you change nothing.
Advanced OpenAI models	Yes — your course's own OpenAI key	Courses whose questions genuinely need heavier reasoning: proof-based mathematics, complex code review, deep analytical work.
Anthropic or xAI models	Yes — a key for that provider	When you specifically want a different provider, including as a teaching point about model differences.

The setting lives under Models and API keys in the teacher dashboard. Keys are per course, and usage on a key is billed to that account by the provider. Advanced models only become selectable for students once a valid key is in place; without one, the course simply runs on the free default.

Background analysis never runs on your key

Cognitive-level classification, retrieval indexing and the other analytic work always run on platform capacity, even in a course that supplies its own key. Your key pays only for what your students' conversations consume.

6. Socratic and debate modes

Two optional modes, enabled per course and organised around topics you define. Both are course features — they are not available in the personal assistants students build for themselves.

Mode	What the assistant does	Works well for
Socratic dialogue	Leads with questions rather than answers, on a topic you set, and keeps pushing for reasons.	Concept-formation work, ethics, interpretation, anything where the student's own articulation is the point.
Debate	Argues a position so that the student has to respond to a real interlocutor rather than to a blank page.	Argumentation practice, preparation for seminar, examining a position the class agrees with too easily.

Using them well

Give them a slot, not the whole course. A single well-framed Socratic session that students know is coming beats leaving the mode on all semester.
Write the topic as a question with a real disagreement in it. "Is civil disobedience ever justified in a functioning democracy?" produces dialogue; "Discuss civil disobedience" produces summaries.
Tell students what the mode is doing. An assistant that suddenly refuses to answer directly reads as broken unless the student knows it is deliberate.
Do not use it for factual lookup. Being questioned about a definition you need in order to finish a problem set is annoying, not educational.

Pre- and post-dialogue assessment around Socratic sessions is available if you want a before/after measure rather than an impression; the method is documented in the Socratic pre/post note (Chinese).

7. Letting students in

Courses are private when created. There are four ways in, and you can use more than one.

Route	How it works	Best for
Join code	Six characters. The student enters it and is in immediately, no approval step.	The normal case. Put it on a slide in week one.
Enrolled list	You paste your roster of student numbers; a listed student who opens the course link joins automatically.	Large courses where you already have the official roster.
Request and approve	A student without the code requests access; you or a teaching assistant approves or declines. Both of you are notified by email.	Late joiners, auditors, students from another programme.
Public course	Anyone with the link joins with no code and no approval.	Open workshops and demonstrations. Not for a normal class.

Practical notes

Do the joining in class. Five minutes with the code on the screen gets nearly everyone in. Assigning it as homework loses a quarter of the class for weeks.
There is a QR code for the same link, which is the fastest route in a lecture hall.
Regenerating the code invalidates the old one immediately. Students already in the course are unaffected; students who have not joined yet need the new code. Regenerate if a code leaks beyond the class, not casually.
Repeated wrong-code attempts are rate limited per account, so a leaked partial code cannot be brute-forced.
The code is visible only to teachers and teaching assistants — students in the course cannot read it out of the interface and pass it on.
You can switch a course between private and public at any time.

8. Coursework through the semester

The tools below are all optional. A perfectly good course uses the assistant and nothing else. Add one thing at a time; adding five in week one is the most reliable way to abandon all of them by week four.

Tool	What it does	What you should know
Quizzes	Drafts questions from your material by type and difficulty; you edit and publish.	Nothing reaches students until you publish it. Treat the draft as a first pass by a capable assistant, not as a finished assessment.
Worksheets	Structured worksheets generated from material or a class recording transcript. Multiple-choice items are marked automatically; open responses are pre-marked for your confirmation.	The pre-marking is a suggestion. You confirm or change every open-response score.
Discussion forum	Course board, with an optional scoring scheme that weights semantic quality and peer response rather than post count.	If you grade participation, use this rather than counting posts — see chapter 10.
Surveys	In-platform surveys with scheduled opening and closing times.	Prefer these to an external form: responses stay inside the same governance and consent arrangement as the rest of the course.
Assignments and gradebook	Submission, marking and a gradebook that the other tools feed into.	Deadlines are entered in your local time and displayed in each student's.
Class recording	Audio, screen or webcam capture during class, with transcripts generated afterwards.	Sharing with students is off unless you turn it on. Tell your class you are recording — that is your institution's rule, not ours, and it applies.
Live classroom tools	Polls, a question wall, word clouds, countdowns.	Useful in the room, forgettable outside it. Try one lecture before planning a course around them.
Announcements and email	Course announcements with attachments, and email to the whole class.	Announcements are the reliable channel; students do not reliably read anything else.
Aida	An independent learning companion that works across a student's whole programme and is deliberately Socratic.	It is the student's, not yours. You do not configure it and you do not see it, and saying so to students is worth doing.

9. Reading the analytics

This is the part of the platform that does not exist elsewhere, and also the part most easily over-read. Both halves of that sentence matter.

Cognitive-level analysis

Every student message is classified against the revised Bloom taxonomy — Remember, Understand, Apply, Analyse, Evaluate, Create — automatically and asynchronously, with no action from you. You see the class distribution, a per-student profile, and how the class average moves across the semester.

Read the trend, not the level. An absolute distribution mostly reflects your discipline and your assessment design. A distribution that shifts after you change something is informative.
Expect a lot of Understand. That is what asking a course assistant looks like, in every course we have seen. It is not a finding about your students.
Method and its limits: the cognitive-level methodology note (Chinese). If you are publishing, read chapter 9 of the research handbook first.

Knowledge graph

Concepts and relations extracted from your uploaded material, with each student's demonstrated coverage overlaid. Its practical use is gap-spotting before revision week: concepts on your map that almost nobody's dialogue ever touches. Requires uploaded material (chapter 4).

Learning portfolios

A per-student view across conversations, assessments and participation. Students can export their own.

What the analytics do not tell you

Message counts are not engagement. A student who asked three excellent questions may have learned more than one who asked forty.
Classification is a model output. It is consistent and useful in aggregate; it is not a ground truth about an individual message, and it should never be the sole basis of a decision about an individual student.
Absence of dialogue is not absence of learning. Some students do not use assistants. That is allowed.

10. Assessment: the one rule

You hold both the grade and the data. The platform imposes exactly one prohibition here and leaves the rest to your professional judgement.

The prohibition: research is never tied to grades

Research participation — signing a research consent, allowing data to be used in a study — is completely separate from course grades. A student who declines, or who withdraws mid-semester, still attends, still uses the assistant, and is still graded exactly as before; only their data is excluded. Offering marks for signing a consent form, or requiring participation to take the course, is a serious research-ethics violation, and no platform feature or wording may imply otherwise. Under the power asymmetry between teacher and student, "voluntary" with a grade attached is not voluntary.

Grading AI use is your call

Requiring students to use the assistant and counting it towards a grade is the same kind of decision as grading attendance or seminar contribution: it belongs to your teaching autonomy, and the platform neither forbids nor requires it. Our role is to give you the data and the quality signals; what you count, and how, is yours.

Design advice: quality signals beat raw counts

Message counts and word counts are trivially inflated, and a scheme built on them teaches students to inflate them. Where you can, grade something the platform has already quality-adjusted:

Forum scoring — weighted credit combined with peer reaction, semantic quality assessment and originality checks, which discounts filler automatically.
Socratic dialogue quality — rated on the substance of contributions rather than their number.
Cognitive-level distribution — closer to intellectual engagement than any volume measure, though see the caveats in chapter 9.

Two practical reminders

Announce it in advance. If AI use counts towards a grade, put the scheme in the syllabus in week one. This is ordinary assessment transparency and it prevents most disputes.
Grading changes the data. If you later want to publish on this course, behaviour produced by a grade incentive is not the same phenomenon as spontaneous use. Say so in your methods, or control for it.

11. Teaching assistants

Teaching assistants share the running of a course. There are two tiers, and the difference is entirely about data.

Role	Can do	Requirement
Teaching assistant	Run the coursework tools, approve join requests, manage members, mark work.	2FA, like every staff account.
Senior teaching assistant	Everything above, plus exporting course data.	2FA, plus signing a confidentiality undertaking before any export is possible.

Grant the senior tier only to people who genuinely need to handle student data — typically a research assistant on a study, not a marker.
Access is per course. A teaching assistant on one course sees nothing of another.
Remove people when the semester ends. Nothing does this for you.

12. When things go wrong

Symptom	Usually because	Fix
Students say the code does not work	The code was regenerated after you circulated it, or they are typing a similar-looking character.	Check the current code in course management and re-share. The code alphabet deliberately omits the characters people confuse — `0`, `1`, `I`, `L` and `O` never appear in a code, so a student who thinks they see an O is looking at a D or a Q.
The assistant ignores your instructions	The prompt is too long, or contains two rules that contradict each other.	Cut it by half and remove the contradiction. Test with the exact question that misbehaved.
Replies never cite your material	Nothing uploaded, still processing, or the PDF has no text layer.	Check the knowledge base list; run OCR on scanned files.
The knowledge graph will not generate	No material uploaded.	Upload material first — it is generated from your content, not from the course title.
Analytics are empty	There is not enough dialogue yet, or the class has not started using the assistant.	Wait. Cognitive-level analysis is asynchronous and needs volume before a distribution means anything.
Verification email never arrives	Spam filtering, or an institutional mail gateway.	Check spam, then resend after the cooldown. If it still fails, email us — we can see whether it was sent.
Students hit an error with your own API key	The key is invalid, expired, or out of quota.	Replace it under Models and API keys. The free default model stays available to students throughout, so the class is not blocked while you sort it out.
You lost your 2FA device	It happens.	Use a recovery code. If you have none, email us from your verified institutional address.

Anything not in this table: [email protected]. Include the course name and what you expected to happen. There is no ticketing system and no guaranteed response time — an accurate description gets a faster answer than an urgent one.

13. First-week checklist

Account created and 2FA enabled, recovery codes stored somewhere else.
Course claimed or created, approval received.
Assistant instructions written — role, course facts, how to answer, boundaries.
Core readings uploaded, and one test question asked yourself to confirm the assistant cites them.
Your AI policy written into the syllabus, including whether use is graded.
Join code on a slide, and five minutes of class time set aside for students to join.
A note in your calendar for week two: read twenty real conversations and revise the instructions.

Planning to publish from the course? Read the research handbook before the semester starts — collecting anything beyond ordinary platform use needs lead time, and that is the one thing that cannot be fixed retrospectively. Need to convince a colleague or a department first? The outreach handbook is written for that conversation.