I spend a lot of time diving deep into YouTube when I want to learn something new. The videos I watch can range from 15 to 30 minutes in length, or perhaps even more. If time was an issue, then I can listen to videos like a podcast while doing other tasks. But remembering information is also an issue. Scrubbing through to a YouTube video is more difficult than referring to written notes.

So, YouTube videos contain helpful information, but it’s time-consuming to learn and recall information from them. Is there something I can use to solve this problem?

Fortunately, there is! I discovered a tool called fabric (through a YouTube video, of course!). It’s an open-source command line tool that queries a large language model of your choice with community-written prompts called patterns.

Let’s see how fabric works, and if I can put it to good use!

Goal

I would like to take a bunch of YouTube videos about a topic, summarize them, and save that output to files I can reference at a later time. My hypothesis is that this will save me time and make it easier for me to solidify information about that topic.

Before getting ahead of ourselves, we’ll explore fabric’s potential with one video first. This will let us really dig into its benefits. We’ll also use some command line applications to assess fabric’s performance.

Let’s get started!

Getting familiar with fabric

Fabric’s quick-start guide is pretty straightforward, and after following the instructions, I installed fabric and was ready to go. Most of the steps in this list are cross-platform, so if you want to follow along, the steps from the link above should have you covered.

fabric is not the only tool that’s included in the installation! yt is a script that gathers YouTube video metadata, including a video’s transcript. The transcript is essential to unlocking fabric’s potential; it will be the input that is sent to fabric for processing.

Configuring fabric and yt

There are some prerequisites to using fabric and yt together that will not be covered in this post. I’ll provide some links in case you want to follow along:

• You need a way to connect to an LLM. In this case, I’ll use an API key from Anthropic to connect to Claude AI models.
  • fabric can also connect to ChatGPT and Google Gemini out of the box, so if you’d like to use those models, find their corresponding API keys instead.
• You also need access to YouTube’s Data API v3. Follow steps 1 through 3 in the “Before you start section” of these docs, and you’ll have your YouTube API key ready to go.

fabric --setup

Now that we have our API keys ready, we can call fabric --setup connect the fabric and yt to our LLMs and YouTube’s APIs respectively. fabric will prompt you for your API keys, so paste them in and press Enter when asked to do so.

➜ fabric --setup

# ...

Please enter your claude API key. If you do not have one, or if you have already entered it, press enter.

Once the setup is complete, we can verify if fabric is connected by listing the available models to use.

➜ fabric --listmodels

# ...
Claude Models:
claude-3-opus-20240229
claude-3-sonnet-20240229
claude-3-haiku-20240307
claude-2.1

Here we can see that Anthropic’s Claude models are available. Great!

Default models

We can now run fabric with a specific model with the -m command.

➜ fabric -m claude-3-haiku-20240307 # ...

However, specifying which model to use with each command gets annoying pretty quickly. Let’s assign a default model that fabric will use.

To assign a default model, use the --setDefaultModel option with the name of the model you want to use. I want to use claude-3-haiku-20240307 by default.

➜ fabric --setDefaultModel claude-3-haiku-20240307

Now we won’t have to specify the model each time we use the application!

Verifying the configuration

Everything should be configured now, so let’s give both commands a test drive and see how they do. Here’s the yt command extracting metadata from a vine boom sound effect video.

➜ yt --metadata https://www.youtube.com/watch?v=Oc7Cin_87H4
{
  "id": "Oc7Cin_87H4",
  "title": "Vine boom sound effect",
  "channel": "Business Goose",
  "published_at": "2021-09-11T19:00:07Z"
}

Here’s a small test for fabric. Our input needs to be piped into the fabric command. The tweet pattern will provide some brief output.

➜ echo "I had deep dish pizza today." | fabric -p tweet
🍕 Indulged in some delicious deep dish pizza today! 🤤 #FoodieLife #ChicagoStyle

The commands are working! We’re now ready to speed up our learning process.

Extracting wisdom

We can combine the yt and fabric commands together to extract digestible information from YouTube. Let’s see how this works.

Suppose I would like to watch this video about 11 LinkedIn Profile tips. This video is 23 minutes and 33 seconds long. I’ll call these tools successful if I can gather what these 11 tips are in this video.

To compare the tool’s effectiveness, I’ll gather the details that I can about the video first. I’ll give myself some advantages, including playing the video at 1.5 times speed and writing notes down. How fast can I gather the 11 tips from the above video?

Gathering these details took me 21 minutes and 21 seconds. That’s about 9.3% faster than just watching the video. Plus, now that I’ve written my notes down, I have them to reference at a later time.

That’s pretty good, but let’s see how well fabric performs in comparison. We’ll put the video link from above into our command:

➜ time yt --transcript https://www.youtube.com/watch?v=8SdbTpOcwd8 | fabric -p extract_wisdom -o fabrics-notes.md

# fabric output ...

real    0m11.632s
user    0m2.905s
sys     0m0.209s

Did you see that? These video notes were generated in 11.6 seconds!! That’s less than one hundredth of the time it took to write notes myself. This is summarized in the table below:

Verifying the output

This looks extremely impressive, but is the output even usable? Let’s verify the generated notes are accurate to what I wrote down above.

To do this, I’ll use grep to search for terms within my notes and see if they appear in fabric’s generated output. Here’s the breakdown of some of the tips from the video above.

➜ grep "connections" my-notes.md fabrics-notes.md

➜ grep "demand" my-notes.md fabrics-notes.md

➜ grep -i "network" my-notes.md fabrics-notes.md

Overall, I’m pretty happy with the notes that were generated. Although the generated notes don’t always go into specifics, there’s no reason why I can’t add them later. To me, the time savings massively outweigh the loss in detail. But, how long will it take to find those specifics if I need them?

It depends on a number of factors, but for this video, I spent 9 minutes 47 seconds skimming through it until I felt comfortable with the details. Combined with the time it took to run fabric and get the notes (11.6 seconds), I spent 9 minutes and 59 seconds getting comprehensive notes from a video that is 23 minutes and 33 seconds long. In other words, it was 57.6% faster for me to use fabric and skim through the video than it was to just watch it.

Once again, this is summarized below.

API Costs

Anthropic

Using Anthropic’s AI models through the API comes at a cost, so I’ll share my findings.

For the video above, I processed 7537 input tokens, and 937 output tokens.

This leads to a total cost of $0.0030555. Put another way, it cost me less than a third of a cent to extract the wisdom from this video.

YouTube API

The default quota limit for YouTube’s Data API is 10,000 “points” per day. Below are estimated costs for listing video captions and metadata.

• captions resource, list method: 50 points
• list method for all other resources: 1 point

It’s clear that downloading one video’s will not have me reach my quota. The cost to me is $0.00.

Time savings

As the old adage goes, “Time is money.” The amount of time spent doing a task has a cost associated with it. Using a previous job as a reference, I’ll pick a wage of $60 per hour. If I was working for a company, this would be the amount of money they’d have to pay me to learn from our sample video.

In other words, by using fabric to generate my notes instead of watching the video, I saved $23.22 worth of time. Skimming through the video to get cleaner notes cost me an additional $5.82.

Conclusion

In this example alone, fabric has already proven to be an excellent tool. I’ve used this tool to speed up note-taking from YouTube videos by 130 times! That sounds wildly impressive because it is! With this tool, it’s trivial to gather important insight from a series of videos about a given topic.

Currently, I’m using the Claude 3 Haiku model, which seems to do well enough. However, if I were to use more powerful models, could I save money by reducing the time needed to double-check the generated notes. My guess is “yes,” because watching video and writing notes is an expensive operation compared to using fabric alone.

Next steps

Why stop at one video, when I can create a script that will convert a series of YouTube videos into notes for me?! This only increases the impressiveness of the savings. I can gather information from a variety of sources, and then review the information to figure out some main points. Perhaps I’ll share my process of creating a bash script, a topic which I’m only moderately familiar.

As I try to process larger and larger videos, running fabric with Claude becomes a bottleneck. Perhaps there are other services that offer faster performance (such as Groq) that can help increase the speed of analysis.

We’ll see what happens next! This is very exciting, so stay tuned for more updates!