🐴 How I used NotebookLM & ChatGPT to summarize my talk

Tips for creating a useful document in 30 minutes or less

Photo by Chris Montgomery on Unsplash

“Thanks for coming to the session today. See you all next time!”

I ended the second session of my AI course. It was already late at night.

I talked a lot. While drinking from my water bottle and looking at the screen, an idea for an experiment came to my mind.

How can I quickly and accurately summarize what I just talked about?

For the students who missed today’s session, it would be great if I could share a detailed takeaway document with them.

So I jotted down what I hoped to achieve with this experiment.

My goal:

Summarize what I just shared into a detailed list of takeaways.

My needs:

• Rely on AI’s help and keep my manual editing to an absolute minimum.

• Make it fast. I don’t have more than 30 minutes for this experiment.

• Make it as close to what I said as possible. I don’t want AI to make things up.

• Include my quotes to make it specific rather than just high-level summary.

---

Why I chose NotebookLM

Given my needs, I decided to use NotebookLM primarily for this task.

I’ve done similar experiments before. I appreciated the factual accuracy of NotebookLM, as it provided sources/references for its summaries. It allowed me to verify where the information came from.

It also let me compile various “sources”—documents, pasted text, links to web pages, and YouTube videos—into a single interface, then I could use chat to ask questions.

On the other hand, my experience with ChatGPT was not great for tasks like this.

Earlier this year, I tested feeding my second book into a custom GPT and then used prompts to retrieve information. However, it often make things up—I had to give many instructions to ensure the results were accurate enough to be useful.

NotebookLM is essentially a RAG product—RAG is a process/framework of improving the quality of LLM-generated responses by grounding the model on the specific data you provided. Here is a good article explaining what RAG is, if you’d like to learn more.

---

Step 1: Gather the raw transcript

I went to Zoom to download the meeting transcript. It was a VTT file.

I took a glance at the transcript. Looks good.

Zoom Transcript

I selected the text and pasted it into a Google Doc.

---

Step 2: Link to NotebookLM

I linked the Google Doc to NotebookLM.

While I could have copied and pasted the text in Step 1, I preferred linking the Google Doc as it gives me more control if I need to make any edits.

For this experiment, I only needed to provide one source. However, for other tasks, I could upload up to 50 files per notebook with NotebookLM, compared to 20 files with GPT.

---

Step 3: Summarize takeaways

I could have just clicked on the "Briefing Doc" button under the "Help me create" section for a quick summary. However, since I wanted a more specific result, I took a different approach.

I provided an instruction in the bottom prompt box. This is what I wrote:

Create a detailed, structured list of takeaways from the document. For each takeaway, include quotes from the document.

As shown below, it generated the takeaways for me. As I prompted, each takeaway included a quote from me. Looks good.

If I click on the reference number, the left panel shows exactly where the takeaway came from. This is super helpful for cross-checking information. It reminded me of Perplexity. In fact, ChatGPT’s new web search feature has integrated something similar too.

That said, I wish the formatting were more organized. Apparently, this is not NotebookLM’s strength, so I decided to switch to ChatGPT for help.

---

Step 4: Improve formatting using ChatGPT

I copied the generated content and pasted it into ChatGPT.

I asked it to improve the format for me and it did the job.

Can you organize the text below into a more structured format? Please ensure that you do not make up any content or change the original wording.

---

The final result

The full document was long, so I’m sharing a portion of it here.

It didn’t capture all the important points I shared in my course, but I was fairly happy with the overall result.

Take a look.

👇👇👇

Understanding Common Prompting Pitfalls

Several issues arise when prompts lack clarity and specificity, hindering the AI’s ability to generate helpful responses:

• Combining Multiple Questions
  Including multiple, distinct questions in a single prompt can lead to disjointed responses that are challenging to parse and refine.

  • "There’s a tendency of a few of you to mix multiple questions together in one ask."

  • "So ideally, you can break them into different questions as opposed to glue different types of questions in one prompt."

• Leading Questions
  Closed-ended prompts (e.g., starting with "Is" or "Are") restrict AI’s responses to simple yes/no answers, limiting exploration.

  • "As a closed-ended question."

  • "And part of it is because we’re asking, you know, is it as opposed to how is it."

Strategies for Crafting Effective Prompts

The session outlines strategies to overcome these pitfalls and improve prompt quality:

• Clarify the Ask
  Use precise and unambiguous language to define the desired outcome.

  • "The ask is the most important thing. You can basically omit all the other things, but the ask has to be clear."

• Contextual Enrichment
  Provide background information, such as user research findings and customer quotes, to enhance relevance.

  • "It would have been better if you can feed ChatGPT with some sort of user research customer quotes."

  • "Yeah, because you’re giving it a mental model to match yours so that it’s actually a good conversation to be having."

• Structured Formatting
  Use elements like numbered lists and bold text for better readability.

  • "Number list. This is also like one layer deeper that you can provide a list of one, two, three, four. Paired with double asteroids."

  • "Bold text itself is a way to differentiate information, to bring another layer of hierarchy to the information."

• Leveraging References
  Guide AI responses by referencing external sources (e.g., books, articles).

  • "If you were to reference the books. Because we are in this discovery phase. So let’s say the Lin start by Erica."

  • "What are the best practices of writing problem statements according to Nielsen Norman Group?"

Navigating AI's Capabilities and Limitations

Understanding AI strengths and limitations is essential for effective use:

• AI as a Creative Partner
  AI excels in generating diverse ideas and framing problems creatively.

  • "And as you all know that ChatGPT is good at this kind of creativity tasks that I often found helpful as opposed to those kind of factual tasks."

• Contextual Dependence for Factual Accuracy
  AI’s factual accuracy depends on the data provided or guiding sources.

  • "But the answer that you get may not be specific enough or factual for your actual study."

• AI Memory and Context Persistence
  AI memory is session-based, and reliance on memory for critical information is unreliable.

  • "Otherwise, it just, as I mentioned, based on there’s some sort of like random algorithm it identifies the memories that they think is important and to store it in it."

(The rest is omitted.)

---

Appendix

For your reference, I called out four other features in the “Notebook guide” panel, ranking them by their helpfulness.

If you're interested, I can write another newsletter exploring more features and use cases of NotebookLM. For now, I’ll keep it brief.

1. The “Notebook guide” in the bottom-right corner is the most helpful CTA in NotebookLM. Whenever you feel lost in NotebookLM’s interface, you can click on it to bring up the panel again.

2. I found the “Briefing Doc” to be the most useful action in the “Help me create” section.

3. The “Summary” section can be helpful when you need a high-level overview.

4. The “Audio Overview” section is where I used to create my AI podcasts. Note that it now has the “Customize” feature where you can give more customized instructions.

---

Thanks for reading!

Until next week,

Xinran

-

P.S. For those of you in the U.S., hope you had a wonderful Thanksgiving holiday last week with family and food. I also wrote about my first Thanksgiving experience on LinkedIn, which I’m grateful for.