Jeonghyeon Lim
Back to notes

Upload the lecture files, and Grok Bot studies them first

What I hand to Grok Bot and NotebookLM when I organize my lecture materials, and what I still do myself.

After each university class, I put the lecture recording into NotebookLM to get a transcript, compare it with the transcript from my phone's recorder, and correct the two into one. Then I put that transcript back into NotebookLM with the slides and study by asking questions. NotebookLM is a notebook tool from Google, and its screen now shows the name Gemini Notebook. It takes hours each week, and the steps are always the same.

Lately I have been handing this routine over to Grok Bot. Once it is all handed over, I will only upload the files and study at the end. I have run it once from start to finish, on one open lecture. Here is who does which step.

Me

Grok Bot

NotebookLM

1Me I upload the recording, any transcript I already have, and the slides to a Drive folder

2Grok Bot Checks the folder every 30 minutes and picks up new files ran

3NotebookLM Turns the recording into a transcript ran

4Grok Bot Compares the two transcripts word by word and merges them into one corrected transcript ran

5NotebookLM Takes the corrected transcript and the slides back in partly ran

6Grok Bot Studies the materials first and writes a lesson map ran

7Me I study by asking questions in NotebookLM

“Ran” marks steps that ran to the end when I tried it on one open lecture. Step 5 is “partly ran”: the materials went into the notebook, but the check that the slides were fully loaded did not pass.

Grok Bot: the one watching the folder

Grok Bot is a bot from xAI that works on its own computer. It can run on a schedule, or start right away from one line in chat. I made a folder in my Google Drive for it to work in and set it to check that folder every 30 minutes.

The working folder in Google DriveInbox where I uploadLessons(lesson name)Corrected transcriptLesson mapIn progress results of each stepShared method the instructions Grok Bot followsArchive
Two ways it startsEvery 30 min Grok Bot checks the folderIf I don't want to wait, I send one line in chatCheck the inbox and process itIt picks up the files, processes them, and reports when it is done
The Grok Bot working folder in Google Drive, with folders for inbox, lessons, shared method, in progress, and archiveThe slides PDF, existing transcript, and audio file uploaded to the inbox folder

The real working folder (top) and the three files of the open lecture uploaded to the inbox (bottom). Folder names are in Korean, and the numbers in front keep them in order.

Once I upload the recording, any transcript I already have, and the slides to the inbox, that is all I have to do. The results are collected in one folder per lesson, and the files I uploaded are left unedited. Grok Bot works from a separate account that can reach only this folder, so it cannot see my other files.

One line sent in the Grok Bot chat asking it to check the inbox and process it, and its reply

Starting it with one line in chat. The chat is in Korean.

The report Grok Bot sent after finishing, in Korean. It says agreement was 96.2% and 324 of 395 differences were left uncertainThe Grok Bot routine setting, checking the inbox at 7 and 37 minutes past every hour

The report Grok Bot sent when it finished (top), and the routine that checks the folder every 30 minutes (bottom). The report, in Korean, tells me what worked and what did not pass. Command output and commands are blurred.

NotebookLM: takes it down, then takes my questions

NotebookLM appears twice. First it turns the recording into a transcript. At the end, the notebook holding the corrected transcript and the slides is where I study by asking questions. Putting the corrected transcript back in failed at first, and went through once the file was uploaded under an English file name.

The notebook with three sources: the corrected transcript, the lesson map, and the slides

The notebook as Grok Bot filled it. The three sources are the corrected transcript, the lesson map, and the slides.

Two transcripts into one

I use two transcripts because I cannot fully trust either one. Side by side, the places where they differ stand out, and those are the places to check again. Grok Bot merges them by three rules.

  • A difference is corrected only when the slides or a repetition within the lecture confirm it.
  • If it cannot be confirmed, it stays marked [uncertain], with both versions written out.
  • Differences in symbols, negation, and numbers are always flagged.

These are three real places from the first run, on one open lecture.

Existing transcriptNotebookLM transcriptCorrected transcript
Do they have to be integers?Do they have to be managers?left as [uncertain]The next line says “So they have to be integers,” but this time it was not corrected, and both versions were kept.
course number 60001course number 6001left as [uncertain]The slides say “6.0001”. Differences in numbers are always flagged.
written differentlykeywordNotebookLM version adoptedThe same word is on the slides, so it was confirmed. This happened in three places.

Here the “existing transcript” is not from my phone. It is the transcript MIT OpenCourseWare provides with the lecture.

The corrected transcript inside the notebook, showing line numbers and uncertain marks

The start of the corrected transcript inside the notebook. Each line has a number, and places that could not be settled keep an uncertain mark (written in Korean) with both versions.

In numbers:

43 minlength of the recording96.2%agreement between the two transcripts395places where they differ
32440274
324 could not be confirmed, left as [uncertain]40 small differences such as spelling or filler words, existing transcript kept27 differences in content, existing transcript kept4 differences in content, NotebookLM version adopted

Agreement means the two transcripts wrote the same words (6,481 words in the existing one, 6,770 in NotebookLM's). Both can be wrong in the same place, so it is not a measure of accuracy against the lecture.

Most of the differences stayed [uncertain]. That is the rule at work: what cannot be confirmed is not corrected. The 395 and 324 are from this first run. After it, I changed the rule so that filler words and repeated words count as small differences. So the corrected transcript, for now, is less a fully fixed text than a text with the places to listen to again marked.

The lesson map: a record of studying first

Last, Grok Bot reads the corrected transcript and the slides and writes a lesson map. Like a map for finding your way, it is a document that notes where to look again in this lesson. The first one, made from the open lecture, has lines like these.

KindWhat the lesson map saysWhere to check
Slides and speech differExplaining the slide example for i in range(7, 10), the lecturer says “7 plus 8 plus 9 … So 15.” 15 is the value after 7 plus 8, and the final sum is never said aloudslide 20, transcript L348
Only said aloudTries 7.1 and 10.3 in range and finds the values have to be integerstranscript L352–L354
Passed without explanationThe slide example that asks “what happens in this program?”slide 22
Emphasized“a computer only does what it is told”transcript L011–L012
Check questionIn for i in range(5, 11, 2), what values does i take? Answer: 5, 7, 9transcript L356–L358, slide 20

Five lines picked from the first lesson map, made from the open lecture. In the real map, each item also carries one of three labels: confirmed, inferred, or unverified.

When I study, I plan to keep this map beside me. I answer the check questions first, and when I get one wrong or feel unsure, I go to the transcript line and slide page it lists. The places where the slides and the speech differ become questions to ask NotebookLM.

Having another model read it again

Besides NotebookLM, I use Gemini in one more place: having another model read what I made. The draft of this note was read separately by Gemini and by GPT. I showed Gemini the rendered page and asked whether the figures were easy to read, and I gave GPT the text. I also had GPT review the script that merges the transcripts, and it pointed out ten problems, including one where the greater-than and less-than signs were treated as the same. The point is to see whether one review finds what another missed.

Where it stands now

The examples in this note are not the real materials from my own classes. So that I can show them without copyright problems, they come from a run on an open lecture: Lecture 2 of MIT OpenCourseWare 6.0001.

Pulling the files from Drive, getting a transcript from NotebookLM, and comparing and merging the two transcripts, writing the lesson map, and putting the results back into NotebookLM each ran once. It was not smooth. The corrected transcript went into the notebook only after the file name was changed to English, and the check that the slides were fully loaded into the notebook did not pass. I have not checked whether this helps with studying.

Open lecture used

Ana Bell, Eric Grimson, and John Guttag. 6.0001 Introduction to Computer Science and Programming in Python, Fall 2016, Lecture 2 “Branching and Iteration.” MIT OpenCourseWare. License: CC BY-NC-SA 4.0.