Turning Lecture Slides Into an AI Teaching Assistant

Neil Liu — Mon, 09 Nov 2026 13:31:42 +0000

Every term, the same three questions arrive in the same week:

“Which formula can we use here?”
“How is this different from the example on slide 12?”
“Is the report supposed to be single-spaced?”

The answers are in the materials. They are just scattered across a folder, a group chat, and the instructor's memory. So students ask a person instead, and the person answers the same thing for the fortieth time.

This is a walkthrough of a no-code workflow that moves that repetition onto a retrieval-based assistant. It is not magic, and the second half of this post is about where it breaks.

The setup: three documents, not thirty

The most common failure is uploading everything at once. Three versions of the same chapter, two scanned handouts, and a slide deck full of equations-as-images will produce answers that contradict each other. The model is not confused; the source material is.

Merge the course into three documents:

Core notes. What was actually taught this term, merged by chapter.
Worked examples. The problems students ask about most, with your solution steps.
Course conventions. Formatting rules, grading criteria, common mistakes.

Supported formats are PDF, Word, Markdown, and plain text. Keep each file under 20 MB, and split long material by chapter rather than uploading the whole course as one file.

Generate and distribute

Upload the three documents. The platform parses them, builds a knowledge structure, and generates an assistant with a persona. The vendor documentation says this usually takes under three minutes, with progress shown in the page.

Write the persona before sharing it. Ten minutes here saves a lot of cleanup:

You are the teaching assistant for this course. Lead with the conclusion, then the evidence. Cite the chapter when you use the notes. If the notes do not cover something, say so instead of inventing an answer. For grading or personal advice, refer the student to the instructor.

Then distribute. You get a link and a QR code; you can also export a standalone page or embed it with an iframe. Put it where students already are: the class group, the course page, the last page of the handout.

Test with three kinds of questions

Do not judge the assistant by whether the prose sounds good. Use three buckets:

In the notes. It should answer correctly and point to the chapter.
Not in the notes. It should admit the gap. If it invents something, tighten the persona and check retrieval.
Requires judgment. It should route to a human, not give a confident verdict.

The third bucket is where most deployments go wrong. An assistant that starts grading essays or advising on course selection is doing damage quietly.

Where it still fails

Accuracy is not guaranteed. Important conclusions still need checking against the original slides. This is stated in the product documentation, and it is true of every retrieval system I have used.
Scanned material is a weak spot. Formulas stored as images, blurry handouts, and tables without headers all degrade retrieval.
Conflicting versions break trust. If two documents disagree, the assistant will pick one and sound confident about it.
Privacy is a design constraint, not a feature. Materials are isolated per account and not used to train public models, but student data should still be minimized before upload.
It does not replace teaching. It replaces repetition.

A two-week pilot

Pick one course you will teach again. Spend an hour merging the materials. Generate the assistant and write the persona. Share it with a single class and collect the questions for two weeks.

At the end you will have a list of materials to fix. That list is more valuable than any benchmark, because it tells you exactly where the course documentation is thin.

The hard part is not the technology. It is the hour of tidying up.

Disclosure: I work on TransWeb. You can try the workflow with one document.

I work on TransWeb, the product used as the example here.