I built an AI tool to summarize YouTube videos and save time

I love watching the world run around in circles :)

We started out by writing text, then went on to recording video/speech, only to realize that text was superior in the first place.

I get the value of videos for introducing new topics to beginners, but apart from the very early stages, i much prefer written information. I can search written information in a way that until very recently has not been possible with video/audio.

And also, my life is too short to view a 45 minute video on something i could probably read and understand in 10 minutes, without the obligatory introduction/self promotion/conclusion typically found in the video.

I think it’s funny you frame this as an information density problem.

The vast majority of people consuming YouTube need moving pictures to keep their attention engaged and sadly enough a voice to offer the illusion of social contact.

If you label yourself as a 'content consumer' you have lost the battle.

This might be optimal for off-the-cuff content, since reading is faster than listening but speaking is faster than typing

Unless stenography ever catches on

I think it depends on what the video is about.

Watching a 3 minute video on how to fix my sink saved me about a hundred bucks.

If it was a 45 minute video, of course that would be useless, but "45 minute video vs 10 minute read" sure feels like a false dichotomy.

It's not a dichotomy, just an example. Dude above is just saying, yes video is good sometimes but for a lot of stuff just skimming a written summary is faster and better.

It's not an either/or thing, just that written summaries of video summaries of written information is just kind of funny :)

I remember the massive annoyance of learning to tie a tie or learning a dance before YouTube. Today I feel extreme annoyance at library documentation for animations and UI components that just includes text descriptions.

Sure it depends on the content of the video. Video is great for learning how to fix your lawn mower but other that it feels more like wasting time, at least if its not for your entertainment but to learn something. Thanks for the link will try it out.

My initial thought is that there are already free extensions that do this by grabbing the YouTube transcript and feeding it to an LLM with a prompt.

I myself wrote my own unpublished browser extension to do this. Your extension name is great, however. This kind of extension has been super handy in seeing if a video is worth watching. It takes seconds to run.

Here's pretty much the prompt I use with a couple small personal things left out:

  Create a summary of a YouTube video using its transcript. You will use the following template:

  """
  ## Summary
  {Multiple sentences summarizing the YouTube video}

  ## Notes
  {Bullet points that summarize the key points or important moments from the video’s transcript with explanations}

  ## Quotes
  {Extract the best sentences from the transcript in a list}
  """

  Transcript:

I’ve been using the free tier of Gemini for exactly this, it can read and summarise YouTube videos by reading transcripts and the auto-generated CCs. Works really well!

Do you use anything for emails?

Not yet, I think Google has a Workspace AI product that does that but we haven’t tried it

Do the person you're sending the email to a favor - skip the LLM and just email the prompt you would have used. If the email was going to be worth a damn the prompt would hve included all the context and information you needed anyway.

Nah, not for sending, for groking; for reading long threads/intent/goals/tasks/conflict etc.

Through which interface?

You can just put this on gemini.google.com and it summarizes/Q&A.

"summarize this youtube video https://www.youtube.com/watch?v=DZprmAVkzRk"

People reinventing wheels at a pace never seen by mankind

I dont see this mentioned enough: https://github.com/danielmiessler/fabric

A location to track prompts for reuse by many people at the very least, specific tools including solving this exact same problem as well.

Even though i am very skeptical of all the ai-bull at the very least it needs some kind of global cache database where people share prompts, input and output examples, and human tagged information on how it works or doesnt work. maybe even linked to real testing on why/if putting in "take a step back and look at the results" type nonsense does anything at all

Some suggestions: I watched the movie and thought the “ask questions” part could be done better.

Going back a problem I often have: I search youtube movies for learning something and click through 4-5 movies until I find a good presenter, not just ads, well explained and good summarized. You might not be able to to a better search algorithm but you could think of on a specific topic summaries a few youtube movies and by this give a good suggestion which to look at/read the summary.

I just don't watch them. It almost always saves me time.

Yeah I mean if an information source is so low signal-to-noise ratio I need to have it summarized, there are probably better sources. And if it's actually high quality, the summarization is lossy.

There's no point to doing this.

The point is figuring out if it is information dense or not. You don't know until you watch it. An LLM can process the entire transcript much faster than you can.

I can tell if a youtube video is worth watching in a few seconds of skipping around and listening to random parts.

Less time, lot less resource intensive than having a whole LLM infer text from it so I can think about it.

Does this work at all on those demonstration videos where the videographer is too shy to add a vocal track, and just demonstrates everything, sometimes with text?

Does it work for videos without auto generated sections like (Intro, etc..) I mean more so like for old videos without Labels or classification?

Could you offer this as a bookmarklet for folks not on chrome?

Looks like you’ve used ScreenRun for your video demo, thanks!