Latest / AI Tools for Practicing Lawyers / AI Builds: Gmail Digest

Transcript

AI Builds: Gmail Digest · AI Tools for Practicing Lawyers

• 0:03
  SPEAKER_01: Welcome to AI Tools for Practicing Lawyers.
• 0:07
  SPEAKER_01: Practical, no-nonsense guidance to help attorneys put AI power to work in their practice right now.
• 0:15
  SPEAKER_00: Over the last few episodes, we've been talking about a massive shift that's happening in the legal tech space and really the AI space.
• 0:24
  SPEAKER_00: We're moving away from the era of prompting and everything you need to know about prompting, although, really, that's still a critical skill to know, but we've elevated that discussion from standalone chat boxes into the era of persistent automated workflows.
• 0:43
  SPEAKER_00: Systems that run in the background, remember how to do a task, and deliver a completed result without you having to babysit them.
• 0:52
  SPEAKER_00: I recently told you the story of how I spent way more than 12 hours building a motion to extend time workflow.
• 1:01
  SPEAKER_00: It was a beast, but it taught me a valuable lesson.
• 1:04
  SPEAKER_00: If using AI is harder or takes longer than your pre-AI workflow, it fails what I call the airport test.
• 1:12
  SPEAKER_00: I don't need to go through the airport test again.
• 1:15
  SPEAKER_00: Suffice to say that if it would have been quicker to drive than to fly, then flying fails the airport test.
• 1:23
  SPEAKER_00: And if it would have been quicker to use a non-tech or a pre-AI workflow than AI, that fails the airport test.
• 1:33
  SPEAKER_00: And I think that's a useful test to keep in mind when you're thinking about adopting technologies into your practice.
• 1:40
  SPEAKER_00: Well, today I want to share a story of triumph.
• 1:44
  SPEAKER_00: I wanted to solve a practical problem for this podcast and for my own personal development.
• 1:51
  SPEAKER_00: You know, I always want to keep on top of AI-related news.
• 1:58
  SPEAKER_00: That's why I'm always looking through LinkedIn to find out the latest information.
• 2:03
  SPEAKER_00: And let me tell you, LinkedIn is a great source of information for AI tech if you start configuring that app properly.
• 2:13
  SPEAKER_00: But I also subscribe to a lot of AI newsletters, uh, the Artificial Lawyer, The Rundown, Legal Tech blogs.
• 2:23
  SPEAKER_00: I pretty much subscribe to all of them.
• 2:26
  SPEAKER_00: My manual process has become really cumbersome.
• 2:31
  SPEAKER_00: And as a result, I've ended up ignoring a lot of the noise.
• 2:37
  SPEAKER_00: For me, it's noise.
• 2:38
  SPEAKER_00: Maybe it's not noise.
• 2:40
  SPEAKER_00: Somebody didn't think it was noise.
• 2:41
  SPEAKER_00: They thought it was important enough to include it in their newsletter.
• 2:45
  SPEAKER_00: But for me, it's become noise because I just have so much inflow regarding these newsletters that I frequently just skip them completely.
• 2:57
  SPEAKER_00: So I decided to build an automated persistent AI producer, a system that will automatically grab my unread newsletters, bundle them together, send them to ChatGPT with a highly specialized producer prompt, format the output beautifully, and deliver a pristine, curated, top five podcast briefing straight to my inbox every morning.
• 3:26
  SPEAKER_00: Now, I'm going to walk through how I built it.
• 3:29
  SPEAKER_00: I'm not going to promise this is going to be fascinating to everybody, but it was an interesting process to go through.
• 3:35
  SPEAKER_00: And if you're considering and acting like this, and you really should, because it's very powerful.
• 3:42
  SPEAKER_00: And by the way, I'm digressing a little bit.
• 4:03
  SPEAKER_00: Anyway, here's how I built it, and some of the hilarious roadblocks I hit along the way.
• 4:09
  SPEAKER_00: So the blueprint was building the core engine.
• 4:12
  SPEAKER_00: Now, if you've watched those YouTube videos promising you can build a custom AI agent in 20 minutes, I'm going to be the first guy to say, yeah, I don't think so.
• 4:21
  SPEAKER_00: And now I'm sure of it.
• 4:22
  SPEAKER_00: But I'll give those creators a tiny bit of credit.
• 4:26
  SPEAKER_00: Once you know the steps, this is actually a workflow you can construct in an hour or so.
• 4:32
  SPEAKER_00: I built this system using make.com, which is a visual automation platform that lets you link different apps together using virtual bubbles on the screen.
• 4:44
  SPEAKER_00: The architecture of my automated producer relies on four steps.
• 4:48
  SPEAKER_00: The first one is the collector, which is a Gmail search.
• 4:52
  SPEAKER_00: I needed a way to isolate my newsletters.
• 4:55
  SPEAKER_00: I didn't want the AI digging through my client emails or my personal spam.
• 5:00
  SPEAKER_00: So I set up a rule in Gmail.
• 5:02
  SPEAKER_00: Whenever a newsletter hits my inbox, Gmail automatically stamps it with a custom label, AI newsletters.
• 5:09
  SPEAKER_00: In make, my first bubble is a Gmail module.
• 5:13
  SPEAKER_00: It's set to search my inbox specifically for emails that are marked unread and carry that AI newsletters label.
• 5:21
  SPEAKER_00: The second step was the bundler.
• 5:24
  SPEAKER_00: They like to call it the text aggregator.
• 5:27
  SPEAKER_00: And this is where people usually get tripped up.
• 5:30
  SPEAKER_00: If Gmail finds five unread newsletters, it wants to send them to ChatGPT one by one.
• 5:37
  SPEAKER_00: That means ChatGPT would give me five separate disjointed summaries.
• 5:41
  SPEAKER_00: I didn't want that.
• 5:42
  SPEAKER_00: I wanted one very pretty, nice digest of the top five stories across all the newsletters so that I'd be able to glance them and say, yes, this is something my listeners might find interesting.
• 5:56
  SPEAKER_00: So to solve that, I used a pretty cool utility in make called a text aggregator.
• 6:04
  SPEAKER_00: It's like a digital trash compactor.
• 6:07
  SPEAKER_00: It takes all those separate emails, extracts the sender, the subject line, and the actual content, and bundles them into one single neat block of text, separated by a clean triple dash.
• 6:20
  SPEAKER_00: Now, instead of five separate chats, we have one unified document ready for my lizard brain.
• 6:27
  SPEAKER_00: So the next step, step three, unleashing the producer, which is the prompt.
• 6:32
  SPEAKER_00: Once the newsletters are bundled, we pass them to OpenAI's GPT-4.
• 6:40
  SPEAKER_00: Now, this is an old model.
• 6:45
  SPEAKER_00: This is far from the most recent model.
• 6:48
  SPEAKER_00: I think they're up to like 5.3 or 5.6, something like that.
• 6:52
  SPEAKER_00: This is 4.4.0.
• 6:54
  SPEAKER_00: But for this process, you don't need the most advanced AI model.
• 7:01
  SPEAKER_00: You can use an older model.
• 7:03
  SPEAKER_00: This reminds me of a story about when my son was first getting into video games.
• 7:10
  SPEAKER_00: I was amazed how expensive they were, but I was also amazed at how inexpensive the prior generation's games had become.
• 7:22
  SPEAKER_00: So you went to PlayStation 3, this just goes to show you how old I am, and there were all the games for$59.
• 7:31
  SPEAKER_00: But if you bought more or less the similar game on the PlayStation 2, you can get it for like$19.
• 7:38
  SPEAKER_00: Of course, doubtful that, you know, a nine-year-old kid is going to want to play the prior generation's games, but I digress.
• 7:49
  SPEAKER_00: But this is the same concept for AI models.
• 7:55
  SPEAKER_00: You don't always need the most current model to do what you want to do.
• 8:01
  SPEAKER_00: So we're in OpenAI's GPT-4.0, and this is where persona prompting comes in.
• 8:09
  SPEAKER_00: I didn't just ask ChatGPT to summarize these emails.
• 8:12
  SPEAKER_00: If you do that, you'll get generic, boring corporate tech speak.
• 8:16
  SPEAKER_00: I don't want to talk about enterprise SAAS updates on this show.
• 8:20
  SPEAKER_00: We're trying to talk about real legal workflows.
• 8:23
  SPEAKER_00: So I set up a system prompt that explicitly defines the AI's role.
• 8:28
  SPEAKER_00: I told it, you are the executive producer and lead researcher for Ronald Dresher's AI Tools for Practicing Lawyers podcast.
• 8:37
  SPEAKER_00: Your job is to read through today's newsletters, filter out the noise, and curate a strict list of the top five most relevant, fascinating stories specifically for an audience of tech forward lawyers.
• 8:50
  SPEAKER_00: Then I gave it strict user instructions.
• 8:53
  SPEAKER_00: For each of the five stories, it has to output the headline, a sharp two-sentence summary of what happened, and the podcast hook.
• 9:02
  SPEAKER_00: One to two bullet points explaining exactly why my listeners will care or how I can talk about this on the next episode.
• 9:10
  SPEAKER_00: And that was the secret sauce.
• 9:13
  SPEAKER_00: By telling the AI why it's reading, to find a hook for a legal podcast, it filters the data through a completely different lens.
• 9:20
  SPEAKER_00: It's no longer just summarizing, it's producing.
• 9:24
  SPEAKER_00: So, of course, you're going to hit roadblocks.
• 9:27
  SPEAKER_00: Roadblock number one, the dreaded wall of text.
• 9:30
  SPEAKER_00: So I mapped my variables, clicked run, and waited by my inbox.
• 9:35
  SPEAKER_00: A minute later, a new email arrived with the subject line, Daily AI Podcast Digest.
• 9:41
  SPEAKER_00: I was really excited.
• 9:42
  SPEAKER_00: I opened it up, and instead of getting that beautiful styled briefing sheet, I was hit with a massive, unintelligible wall of text.
• 9:51
  SPEAKER_00: When I say unintelligible, I mean obviously I could read the text, but it was not going to be something that I would be able to just kind of look through at a glance and say, oh yeah, that's the story I want to think or explore a little bit more.
• 10:06
  SPEAKER_00: It was just a single continuous run-on paragraph.
• 10:09
  SPEAKER_00: So in my final Gmail step, the body type was set to raw HTML.
• 10:16
  SPEAKER_00: This is this is how I fixed the wall of text problem.
• 10:19
  SPEAKER_00: The problem is HTML is literal.
• 10:22
  SPEAKER_00: If you don't explicitly tell it to create a space using code, it just runs everything together.
• 10:28
  SPEAKER_00: My first thought was, okay, I'll just go back into ChatGPT and tell it to write an HTML instead.
• 10:34
  SPEAKER_00: So I didn't try to force ChatGPT to become a web designer.
• 10:38
  SPEAKER_00: Instead, I used a beautiful, elegant tool right inside make.com, a markdown to HTML converter.
• 10:45
  SPEAKER_00: I dropped this tiny converter bubble right between OpenAI and Gmail.
• 10:50
  SPEAKER_00: Now ChatGPT can output its natural, comfortable markdown.
• 10:55
  SPEAKER_00: The converter instantly translates it into pristine HTML code and feeds that to Gmail.
• 11:01
  SPEAKER_00: And the result was absolute perfection.
• 11:04
  SPEAKER_00: Clean spacing, bold headers, and beautiful readable bullet points.
• 11:09
  SPEAKER_00: I am going to put links to both of my outputs in the show notes so you can see a before and after.
• 11:18
  SPEAKER_00: And you know, it's really critical that I confess that I didn't dream this up by myself.
• 11:26
  SPEAKER_00: I'm not capable of actually creating this workflow.
• 11:32
  SPEAKER_00: I had Gemini walk me through every step of it.
• 11:35
  SPEAKER_00: Why Gemini and not some other AI tool?
• 11:42
  SPEAKER_00: Because I knew that I was going to be doing a lot of this in Gmail.
• 11:46
  SPEAKER_00: And I wanted a model that I felt was going to know Gmail inside out.
• 11:54
  SPEAKER_00: So now, every single morning, while I'm having my coffee, the silent executive producer runs in the background, sweeps my labeled emails, aggregates them, runs the analysis, translates the formatting, and drops a pristine podcast prep sheet directly into my inbox.
• 12:15
  SPEAKER_00: And let me tell you, the output is gold.
• 12:18
  SPEAKER_00: Just this morning it flagged a trade secret lawsuit between Apple and OpenAI over a Johnny Ive hardware device, upending the usual IP litigation conversation.
• 12:29
  SPEAKER_00: It flagged Meta facing a lawsuit over AI biased layoff algorithms and others.
• 12:37
  SPEAKER_00: And these are highly sophisticated targeted show ideas that would have taken me an hour of manual digging to compile and summarize.
• 12:46
  SPEAKER_00: Now it takes me zero minutes.
• 12:49
  SPEAKER_00: So I invite you to consider how a workflow like this might help you in your own practice.
• 13:00
  SPEAKER_00: Maybe you don't want to be chained to your email clients.
• 13:04
  SPEAKER_00: Maybe you want to have emails digested for you every morning or at the end of the day.
• 13:11
  SPEAKER_00: So you could just kind of read through a digest instead of reading through all the emails, you can filter out the spam within this process.
• 13:23
  SPEAKER_00: I imagine that there are countless use cases that would really elevate an attorney's practice if they took the really fairly brief amount of time that it takes to set this up.
• 13:39
  SPEAKER_00: My name is Ron Drescher.
• 13:41
  SPEAKER_00: This is AI Tools for Practicing Lawyers, and we'll see you next time.
• 13:47
  SPEAKER_01: That's it for today's episode of AI Tools for Practicing Lawyers.
• 13:51
  SPEAKER_01: Thanks for listening.
• 13:53
  SPEAKER_01: We'll see you next time.