Subscribe
Share
Share
Embed
Bots at Work is the new season from Fringe Legal, which explores how AI is changing the way work gets done, with a focus on real-world impact over hype. It looks at how operators, builders, and leaders are using AI to reshape workflows, decision-making, and business models, especially in professional services like law. The show focuses on practical insights, emerging patterns, and honest conversations about what works, what doesn’t, and what comes next as intelligence becomes cheaper and more embedded in everyday work.
Ab (00:31)
Well, Mike, welcome to the pod. So glad to have you here. And I'm excited for the conversation today.
Mike Brown (00:38)
Stoked to be here, Ab. Thank you so much, man.
Ab (00:40)
Yeah, so we met randomly only a few weeks ago, of all places at a anthropic meetup here in Orange County. And I was excited to meet another lawyer there for what seemed and I think for most people would have assumed it to be a very tech focused event. And you were up there giving a demo. So you have a really cool story where you
basically beat out something like 500 engineers at a hackathon recently, which kickstarted your, I guess, your software development and your company formation journey before you were a practicing lawyer. So can you mind just giving us a bit of your origin story where it all started?
Mike Brown (01:23)
Yeah, so I was a lawyer, still am a lawyer, got a little bored, running my own personal injury firm, started getting really good at prompting. And then with my cousin, we might be able to go into that, but he's a serial entrepreneur. He convinced me to try to connect my prompts into something we could sell to other law firms. Then through that, we hired a professional coder and he did a great job with it, but then we needed to make some changes and we were basically out of money and not making money.
like most startups. And then it became a necessity thing of, okay, what can we do? And that was right when cursor was like just coming to the fold. These AI tools were getting good enough where you could just kind of tell it what you need. And then either copy paste or cursor would just do it for you. And it was kind of crazy. And then, so I learned coding that way. I still to this day do not read code. now I tell Claude. And then,
I ended up at this, I applied for this hackathon. was 13,000 people that applied for the Anthropic Clod Opus 4.6 Builds Hackathon. And I didn't think much of it because I was just like, all right, yeah, I'll apply, but I'm not getting in. And then next thing I know, I get in. And then next thing I know, I end up winning the thing. And I came with $50,000 for Anthropic credits, which has been really helpful. And it's honestly, it's been a life changer. It was kind of like,
The way my mom explains it, it's like I won the American Idol of Coding
Ab (02:48)
That's super cool, actually. And I don't know some of that. it sounds like we'll rewind a little bit. And for a lot of the viewers, listeners, cursor is this place where you can code. It's usually for developers. It's called an IDE, Integrated Development Environment. And you basically write code in there. It's what traditional coding is, but you can access all of these other different models. I also use it.
It seems daunting because if you're not a developer, if you're a lawyer, you are going into the weeds a little bit. And I wrote a post about how it's become so normalized for so many knowledge workers to now start their day and continue their day from the terminal. I don't think this would have been a thing even a year ago where for most people asking them, hey, can you do this in your terminal?
they wouldn't know what the terminal was, let alone, you know, have it open and have lots of instances of open day in, day out. But that's the, that's the beauty of Claude code, I guess. But it seems like you've been down this path of sort of exploring and learning large language models and GenAI based tools. What other than boredom, because that's, well, that's a good motivator. I don't believe that's the whole story.
What was the big driver for you? What actually got you started in the first place? Do you think you've been quite a entrepreneurial or exploratory or creative person otherwise?
Mike Brown (04:14)
Well, so before law school, was actually a visual effects artist and a little bit of a reality TV producer. So I've always had this creative itch. you know, and I kind of got burnt out of the Hollywood life. And that's why I went to law school. And I actually thought in law school, I was going to go back to entertainment law. And then while I was doing that, I kind of was like, wait, the thing I hate about Hollywood is the kissing butt. And so through that, you know, it's just kind of like, it's like, all right, I guess I'm not going back there.
Ab (04:39)
Right.
Mike Brown (04:44)
got into PI, but I still always had this creative itch I had to go do. it was kind of incredible, like with the launch of ChatGPT, and then all of a sudden it's just like, there's this side of me, that creative side, that I couldn't, I wanted to keep it tempered down, because I was making decent money on, well, working for a PI firm, then I was making that transition to my own, which is, you're gonna be pretty broke for a little bit, that six to nine month period, waiting for that first settlement. And I was just,
Ab (05:07)
Sure. Yeah.
Mike Brown (05:10)
I was just getting through there and that's when I started realizing I was getting pretty good at the AI stuff.
Ab (05:15)
Yeah. And if you don't mind, because I think there's a lot of people out there, a lot of lawyers out there, and generally knowledge workers who may find themselves in that boat. And certainly there's a lot of stories, including yours, where people might feel, and hopefully they do, they feel encouraged that, yes, I can do it too. But one of the biggest and the most common things that comes up is, I want to do this, but there is a learning curve.
in anything, right? Even if it is just open up and start prompting, there is a learning curve, there is a comfort curve. How did you find the time because there's only X amount of hours in a day. But how did you sort of put the time aside for that to learn and figure out, okay, I should use cursor, I should do this. Did you lean on other people? Was it just, you know, just experimenting on small projects? What was your start?
Mike Brown (06:12)
well, look, I have a natural curiosity, I would say, that does separate me from a fair amount of lawyers. But it was a necessity aspect of, once we started this thing called OnBreeze, which is basically Granola, which is an AI note taker, but it was for lawyers, right? To be fair to myself, I started it before Granola was a thing. we came out with, but.
Ab (06:29)
Hmm.
Yeah.
Mike Brown (06:35)
to give myself credit, but Gorilla is amazing. These other AI note takers they're all great. Mine was a phone line. But it was kind of out of necessity because it was just, we needed to be able to scale this thing to handle more than four or five people. And that was the start of it. What I will say for lawyers looking to get started, right? Look,
Ab (06:47)
Yeah.
Mike Brown (06:54)
If you dedicate yourself to one week of I have Claude, have ChatGPT or Codex specifically, get yourself into one of these agents. If you tell yourself for one week, if something takes me 10 minutes to write this email normally, I'm gonna take 15 minutes and force the AI to do it. And you just dedicate yourself for that single week. Within a week, you will now be seeing the compounding of the skills. And before you know, you're gonna start building like Skills.md, capital S.
Ab (07:19)
Yeah, right.
Mike Brown (07:20)
That's
when it starts paying off because rather than being like explaining your, the way your style of your email works, it's like, you're, can basically just set up cron jobs where it looks at your emails and then it's like, it writes the drafts for you beforehand. And now you are saving yourself time. You are basically creating value for your clients without having to do that necessarily. I actually truly believe lawyers are well positioned to be great.
Ab (07:32)
Yeah.
Mike Brown (07:43)
at this new era of AI agents. And it's just, it's gonna require people to have a little curiosity and be willing to go outside their comfort zone.
Ab (07:50)
Yeah, no, completely agree with you. And I do think that that one week, I think even a weekend, you know, just put aside one weekend, I know it's hard. I know there's always things to do. But the return on investment is just immense. It's unreal. And I it doesn't take long. That's the thing, especially for most people, when I talk to them, I show them one small thing. you know, at this point, I've not opened
Claude, I've not opened cursor, I've not even opened cowork is just in the standard application. And I think even in there, people just are not using anything, right? Anything more than just using it as a research tool. So just go that 1 % beyond, know, try a deep research to understand and go really deep and you get, you know, like a BCG, Bain or, you know, a McKinsey level researcher at your fingertips.
my God, what? And you can just, you don't have to do much. You just ask a couple of questions your way. And 30 minutes later, it's looked through hundreds of sources for you and presented you something that's outstanding. think people will start to see a return on that. The other thing which I will just call, yeah, yeah, yeah, Yeah.
Mike Brown (08:50)
Yeah.
see real quick to go with that app. ⁓ You know,
it's kind of like, okay, so a year ago, the models were what they were, right? it's not a linear curve right here. It's exponential how much better they are. But the problem is a lot of people are just interacting with it within chat GPT, which is essentially just Google on steroids, the way I look at it, right? It's just search with a little bit more fun to it. What you really need to do is try to get it to do something for you. Create a PowerPoint, create a visual.
Ab (09:14)
Yes.
Mike Brown (09:31)
You know, one of my favorite ones to tell people, you know, lawyers don't love this one, but I tell them, just create a video game this weekend, you know? And so if you started doing this a year ago and you asked to do a video game, was actually, it's kind of fun because you had a problem solve, think through it and it was steps and it would take a couple hours, but it was fun. Right now, these new models today, Opus 4.7 came out. You can literally be like, create asteroids, that video game, but like make me, put me in a first person shooter form and it will do that.
Ab (09:36)
Yeah.
Mike Brown (09:59)
15 minutes later, like you know you don't have to think much about it, but it's really fun to have it do things.
Ab (10:04)
Yeah, I agree. I think having fun whilst doing some of these things, especially during the learning phase, it's important because I think, we'll talk a lot about this, to be realistic, the more complex that your task becomes, the more you do have to think, right? And the less of an expectation you should have, understandably that, hey, this is going to take me five minutes, right? If you're there trying to say, I want to, we're recording this on Riverside.
I want to create a Riverside alternative. This is not a five, 10 minute thing. There's a lot that goes into that. But there are so many other things where you can get outside for 10. yeah, and yeah, yeah.
Mike Brown (10:43)
Well, going with that, right? It's
the complex task, right? People are thinking, you know, they think of it as like, the AI is just not good enough because they expect the AI to do everything. And what you got to realize, this is an incredibly smart partner that you're gonna be working with that is like, you got to guide it on. And it takes skills that are learned every day by using it. So if you were to go build a Riverside like this, there's actually a lot of components that if you just tell it to go do Riverside, it could kind of do some of stuff, but it's gonna miss like the client side uploading.
Ab (10:51)
Yeah. Yeah.
us.
Yeah.
Yeah.
Mike Brown (11:12)
going to miss the compression
Ab (11:12)
Yeah.
Mike Brown (11:13)
stuff because those are things that you need to understand yourself. And if you ask it to do it then at that point it will, but those are skills you built up over time and multi-step processing. But that's actually where I think lawyers who can, who every day.
Ab (11:16)
Yes. Yeah. Yeah.
I agree.
They're so good at deconstructing things and breaking it down into their sort of atomic levels. I agree with you. know, I was having a conversation with someone yesterday and I told them, same thing to you. I'm not a developer, but I've been vibe coding before it was called vibe coding for a long time. But my, I think my superpower in this world has become, I am so good at reading documentation, both technical and otherwise, right? And
to be able to create a PRD, to be able to create a spec, to be able to cleanly plan out what do I want to happen. In fact, sometimes I have to stop myself because I want to give the language model the creative control and not tell it exactly what I want. I want there to be a chance of randomness. But I think lawyers naturally, because you spend so much time learning how to do this.
Right? You get so good at this and you can go through and when I'm reviewing API documentation, I can tell you I don't understand any of the code, but I absolutely understand the English and I'm able to understand the capability and it's just putting the puzzle together.
Mike Brown (12:37)
And going with that, man, what's so funny to me about that is some developer was when I was explaining, I read API documentation. He was like, it's so boring. Why do you do that? I'm like, like, well, it's actually more it's more interesting than what we read for three years at law school. Like lawyers, we already understand how to go find the fun in some of the most driest legal documents in the world. So and that's what API documentation, all that. But we know how to go into realms that we don't understand and go find the relevant information really well, because we've all been trained that way. And that's kind of this vibe code.
Right? So it's not math. It's reasoning.
Ab (13:11)
Exactly, yeah.
So, okay, great. So yeah, I think for anyone listening, if you're not already itching to go and do something, learn something, go create something today, I promise you, you're gonna have a lot of fun doing it. There's lots of places and you can hit either one of us on LinkedIn. We'll include our details in the show notes and we're happy to guide you in different ways. Okay, let's go back to your story. let's talk to me about the hackathon. So you applied for it, the 4.6 build hackathon.
on a whim. Of course, we know what happened at the end. But what was the idea and where did that come from? And how did you develop that as part of the hackathon?
Mike Brown (13:46)
Yeah.
Yeah, so what got me on this one was funny is the idea I applied for. I'd already come up with a name, so it's actually called Crossbeam. And it's basically California permit system is a nightmare. Any builder or even on the city side or if you're a homeowner or anything like that, you've been through it. You understand it's like the permit gets submitted to the city. They have a queue because they are backlogged with it. Almost any given city in California. It takes a long time for them to get back to you. And then you're going to have things that
the builder miss, right? On this blueprint. So then the builder then has to go do it. And then it creates this whole cycle of two to three correction cycles. And that's a good day, right? So where I came across this idea of how can we use AI, specifically Claude code to help with this is ⁓ I was building a demand letter generator like everyone else in legal tech. I built it in November and I did it in a way so that it was cheaper than everything else out there. It's pretty dang good in my own book. Anyways, besides the point, not trying to sell that.
But so that that demand letter software, I tested it with zip files that contained up to 1500 legal docs and medical imagery, all different types of files inside of a single zip file. And then someone my buddy, And he was telling me he's a builder of ADU's. His name's Cameron, ADU West Coast. They're excellent if you want an ADU, great value. ⁓ Shout out Cameron. And he was telling me about this permit process. And I was like,
Ab (15:07)
Shout out to Cameron, yeah?
Mike Brown (15:11)
there's no way like AI can fix this. Now, so he passed me one of his blueprints, I sent it through Claude Code and it actually aired out Claude Code and this was in November. And that kind of really fascinated me because basically to this point, the 97 % of things I had thrown at it, it could handle. And so then I dove into it a little bit more.
Ab (15:20)
Mmm.
Hmm.
Mike Brown (15:34)
And it found out like, so there's not to go too technical, but there was vision requirements within the API that would cause it to tear out. These blueprints are massive. They're meant to be printed on a piece of paper the size of a table, right? So there's stuff like that. Then if you go not to go too technical again, sorry, I'm sorry say that. It's the context limits at the time, especially, were 200,000 tokens. Now,
Ab (15:51)
No, we can't. It's fine.
Mike Brown (15:57)
As a lawyer, like that's where I recommend people to start understanding is context engineering, how you handle multi-turn conversations with your AI to manage token context. That's when people say they have hallucination issues and stuff like that. It's like, no, no, that's because your context ran out in that chat GPT conversation because they have limits to it. So then the AI gets like an anxiety. So it just kind of makes things up to finish. And the way you get around that is basically either there's different strategies you can do. And it's basically
taking the finishing end product into another kind of conversation window to continue and fact check and you just kind of keep looping like that. Or you can use in Claude code, there's sub agents. Sub agents are basically mini chat GPT conversations under the hood. And so that's how you can kind of handle context stuff. But anyways, back in November that fascinated me. was just an idea that was like, wow, this didn't work, whatever. When the thing came around, it was like, it asked the question of what would you like to pitch? And I was like,
Ab (16:49)
Mmm.
Mike Brown (16:56)
I don't even think I'm going to get in. I'm like, okay, well, I've been sitting on this idea for a little bit. My buddy's ADU permit thing. Like, all right, we'll give it a shot. So I get in and when I got into a shock, I got in because I'm up against huge, like people, engineers from nVidia and Apple were in this thing. It was crazy. And, and then I was also like, shoot, like I'm doing something that I know is actually kind of impossible. But I was like, all right, you know what? It's a week of my life. We'll give it a shot. ⁓
Ab (17:00)
Great. Yeah.
Right, right.
Right.
Yeah.
Mike Brown (17:23)
And so I came up with, combining with my VFX background, I came up with some nifty kind of cheat codes, I guess you could say, to be able to force the images in and break it up into pieces and then be able to have it use the vision. And the big change was that hackathon was for Opus 4.6, that had just came out and prior was Opus 4.5. 4.5 would break down and then 4.6, if you gave it the proper image breakdowns, ⁓ it actually started working. And I remember seeing that
Ab (17:29)
Mmm.
Mm-hmm. Yeah.
Hmm
Yeah.
Mike Brown (17:51)
If the hackathon started on Tuesday, I remember Wednesday night throwing through locally the initial permit thing. And then it was like, wait, this didn't error out. And it was like, it was that, yeah. And so that was really cool.
Ab (17:59)
Yeah. It's not erroring. Yeah. Yeah. Yeah. I think, and
I think it's so cool though. And I think, again, these are just some things that people should think about because things are moving so quickly. And I can't remember who said that, whether it was Dario or Sam Altman, which for the builders, you should be developing at the, at the edge at the frontier.
where you are getting up against the limitations. I wish the model was smart enough because probably in one to six months, the model will get smart enough. And I think you and I spoke about some things that I've been building. I've been doing some stuff with voice for sales and B2B sales team. And for the longest time, that was my limitation where I was just like, can do this, but the voice models just aren't there yet for what I want it to be, at least at the
Mike Brown (18:52)
Mm-hmm.
Ab (18:54)
quality I want it to be. And then suddenly an update drops and you're like, well, okay, now everything gets there, right? But all of the time that you've spent, what you call cheat code, I call creative problem solving for you. It pays off because you figured out a efficient way to be able to get around the problem. And now, because the base intelligence is higher, you can now do all of that so much better, presumably.
Mike Brown (19:08)
Hehehehe
Wanted to go with that today being model drop day is this my new life as like a developer, guess you could say, I Opus 4.7 came out today. Massive, massive, and we won't go into the 1586 in the past to 2564 that came out today, but it I mean, look, it's I put all this work into this thing and I got to work and Boris Cherny, the creator of Claude Code, talks about almost exactly what you brought up, which is don't build for the models today. Build for the models in six months.
Ab (19:25)
Yeah.
Yeah, they were big vision updates in that, right? Yeah.
Yeah.
Mmm. Mmm.
Mike Brown (19:51)
And literally this morning I got to see a payoff because 4.7 came out and I've already got a 7 % performance boost by just changing the model. Now, when we finish up with this podcast, what Claude is actually, well, it's working on it right now. When I press go on the plan mode right before this podcast, but it's redoing my vision pipeline to match the newest vision procedure parts. And so I think we're gonna get another 7 to 10%. And then.
Ab (20:07)
Yeah, right.
Nice.
Mike Brown (20:18)
But I mean, imagine if you go build something
Ab (20:19)
Yeah.
Mike Brown (20:20)
and then in a single day, they're like, you want to be 15 to 20 % better? just, and I'm not doing much. Yeah. It's a no brainer. Yeah. Yeah.
Ab (20:24)
Right, exactly. Well, what a no-brainer. Yeah, exactly. Yeah. Can
we take a quick aside as we're just talking about model drop days? Are you using the, have you tried the new ultra plan mode yet? are you using this? are? OK, what do you think?
Mike Brown (20:36)
Yes, yeah.
UltraPlan is great because it's kind of like an async workflow in the sense because it works in the cloud. ⁓ so Plan Mode is, it's a great technique to basically prime your agent for something it's about to do. You give it kind of like some ideas, what you're thinking about, and then it goes in, looks through your code base and looks for it. UltraPlan Mode does it in the cloud through your GitHub.
Ab (20:43)
Mm-hmm. Yeah.
Mm-hmm.
Mike Brown (21:01)
So the idea there is it should look through more. It does take longer. I love it for ⁓ future features I'm gonna go build out. Now, I was hoping for like, to me, I would've been cool if it spent an hour or two hours and it went through my whole code base. It still has some holes, but it is better planning in that direction. What do you think of it?
Ab (21:09)
Yes.
OK.
So I used it. So I think it was in Cloud Code before 4.7 dropped. So I used it a couple of times last week. So it erred out on me half the time, which was not good. I did have my environment set up on GitHub, but yeah, I did, which is why it was surprising and frustrating that it erred out half the time. ⁓
Mike Brown (21:28)
Yes.
You need to your environment set up on Gitem. you did? Okay, sorry.
Okay.
Ab (21:49)
So, but I think the plan it created was good. It was actually a much better plan and the advantage were again, like we are being technical, But I think the advantage is A, you have a better starting point, which also means that it's more token efficient. And when you start having, you your, your limits both from a context point of view and also
your usage limits or you're paying for API credits or costs, you have to think about those things. But the plan it created was significantly better because in cursor, I use plan mode almost every single time I'm doing a feature development or anything big, just because I want to read again, know, lawyer in me, I want to read the plan for how are you approaching that? Because sometimes
Again, I don't understand the code and all the random things that it puts in there that we're to set up this or that. I'm not really interested in that, but I'm more interested in the approach. Because sometimes you can identify, actually know you misunderstood what I was saying, probably because I wasn't clear enough, or it just misunderstood. And you can correct it then rather than wait for two hours whilst it's doing something.
So overall, I'm excited for it. hoping, I haven't tried it since launch today. I've got a bunch of projects that sitting. I'll be working on this evening, but yeah, I'm looking forward to it. Yeah, I think so, I think so.
Mike Brown (23:10)
He'll knock him out, man. But
to go with that too, man, I fully agree, right? To me, I've been explaining it like, I don't remember the exact Abe Lincoln quote, but I think he said, if you give him an hour to cut down a tree, he'll spend 50 minutes sharpening his ax. And that's the same thing with these AIs, right? You need to plan it out like very deeply because it can get any plan done really quickly.
Ab (23:26)
Yes. Yeah.
Mike Brown (23:37)
But the problem is in a week, two weeks, a month from now, I'm eight weeks into building Crossbeam out since the hackathon, right? The things I cut corners on earlier, like some days it comes and bites me in the butt. And so it's better to spend the time upfront and cheaper ultimately to spend a lot of time planning. I think that's also where lawyers are really good, because it gives you that plan. Now, one thing I want to, like a technique I recommend to a lot of people, you take that plan, you bring it over to another model, and then
Ab (23:37)
Yes.
Hmm.
Mike Brown (24:04)
that other model, like give me the broad strokes of what this would produce. I call that adversarial prompting. And because it gives you it like you have this plan in your head that you want it to go do this feature or like a whole program or whatever. And if you go to another model and it tells you, it will tell you what it thinks it's about to create. And if you see, it will basically describe like, here's your holes. It won't say, here's your holes here. You'll just see, no, that's not what I meant right there or whatever. And then you take that info back to the original planner.
Ab (24:05)
Yes.
Yeah.
Hmm.
Yeah.
Mike Brown (24:32)
you just keep working on it you loop back and forth between the two agents. yeah, that's a planning mode is so important. it's yeah, that's when people think they're just gonna go their AI and be like, hey, make ⁓ like make research report go like, nope.
Ab (24:34)
Yeah. Yeah.
Yeah.
Right. Yeah, exactly.
But I think it is, you know, overall, I feel like generally the analogy is that it's like working with a very good executive assistant that you've just hired, right? That person has a lot of raw skills, but they just don't understand what you want yet, how your company works, how you like to work, all your SOPs and everything else. And as you start working on a project, and as you start working with
know, Claude ChatGPT, whatever the model of choice is, which I think for both of us is probably all of them in some capacity or other, you start feeding it that information, right? You have your skills.md files. You also have your project brief. You have everything else. And then you have, this is what I want. But you have to give it that context to be able to expect the result that you would expect from.
someone more. And I mean, if nothing else, imagine you have access to probably quite a senior engineer at, you know, for free or for $20 a month. It's wild. It's absolutely wild.
Mike Brown (25:52)
It's as I said,
do you use voice mode?
Ab (25:59)
course. but okay, so well, let's talk about that a little bit. So we both use a software called Wisprflow Right? I think when the Anthropic meetup, it seemed like probably something like 80 % of the people there use Wisprflow So shout out Wisprflow. Feel free to reach out if you'd like to sponsor.
Mike Brown (26:03)
Yep.
Yep, please sit.
Ab (26:25)
or give us
free credits, exactly. So I use Wisprflow for most things, certainly when I'm on my desktop. But do you mean that, or do you mean voice mode that's built into Cloud Code, or actually talking to Cloud or Chat GPT OK. Yeah.
Mike Brown (26:39)
I use Whisper Flow all day, every day. ⁓
So they give out an email saying ⁓ what place you are in your cohort. And so I've always been second place in that email. And during the hackathon, actually met the guy that was first place. I was looking at it a couple of days ago and I was 2.3 million words dictated. But so I use it. It helps me think. I'm like verbal and it really, I think through things. I use it all the time.
Ab (26:47)
Yes. I need to look now. Yeah. ⁓ cool.
Wow, okay, I'm not there. Okay, I'm not there, but yeah, that's impressive. Yes.
Mike Brown (27:09)
But so in general, don't, I don't, when I'm recommending to people just getting into it, you don't need to go straight to whisper flow that every model provider chat GPT has one anthropic has one start there, start to learn how to use it. But it is the ultimate cheat code when you're working with these LLMs because, and it goes back to what you're talking about with the context thing. You know, it is, it is, someone will figure out the memory thing to know who you are, but each time you chart up a new chat, you need to kind of give it that context of like,
Ab (27:22)
Mmm.
Hmm. Yeah. Yeah.
Mike Brown (27:38)
who you are, what we're doing here and why. And the why is the most important part to me, that intent. When you make sure that it understands why you're doing something and why you're doing it in a certain way, it becomes such better output. So that's why I go into the law school method of IRAC. I use just literally Wisprflow and I talk for five minutes straight and it's just, I lay out the issue, I give it the rules, here's where the files are, pay attention to this, the docs, whatever.
Ab (27:52)
Hmm.
Yeah.
Mike Brown (28:07)
And then that analysis section that I just think of as a brain dump. I'm just going to talk for three to four minutes sometimes. And then conclusion is the very end, the science is showing that it beheads most attention to the front, that are the first couple of lines, and then the last couple of lines. And you just remind it, hey man, this is what we're doing. Follow the rules, do it, go. And then you press enter.
Ab (28:11)
Hmm.
Yes.
No,
I agree. I think, again, couple of years ago, maybe a year ago, we generally you have to be focused on prompt engineering, you have to think about, know, act like this, do this, you're thinking a lot, the models are super intelligent now. And now it's just a matter of just just talk to your stream of consciousness is fine. The model will figure out what you what you want, as long as you're at least specific about I want to do this, right? I think they do a pretty good job.
Mike Brown (28:52)
Well, especially because if you adopt this like very strict planning kind of architecture, know, that stream of conscious prompting that you're doing is leading to a plan. And that plan ultimately is the prompt engineering that people like became so obsessed with. But it's just a lot easier and faster to do it. Now I pay attention to context engineering matters a lot nowadays. ⁓ That's a whole other beast, I call it look up a ripper.
Ab (28:56)
Yes.
Mm.
Yes.
Yes, absolutely. Yeah. That's a whole other beast. Yeah, I think we've sort
of touched on enough technical topics for today, though. ⁓ So, okay, let's go back to a couple more things that will be relevant to this audience, which is, you you're working with large documents. That's very relevant. In your case, you're dealing with images. It seems like both your creative problem-solving and model improvements is helping there.
Mike Brown (29:22)
Hahaha
Ab (29:42)
How are you thinking about, mean, you you're a lawyer, you're still working with, well, I guess like city councils and governments and stuff. the confidentiality dealing and handling sensitive information, there's PII in there and all this other stuff. How are you thinking about that?
Mike Brown (29:58)
Yeah, I mean, the PII just had to go through this with an insurance company. Luckily, it's a lot less PII than what we deal with as lawyers. So it actually felt a little bit easier. Now, okay, if you're going to go out there and vibe code an app, yes, in general, you should learn some security stuff. You should read up on SOC 2. It's really not that complicated when it comes to that.
Ab (30:02)
Mmm.
sure.
Mike Brown (30:19)
But on the business side of things, yes, I'm dealing with cities and so I got a pilot program with Buena Park. I'm very excited about and then some other cities I'm talking to. so, you know, and that's kind of where the understanding the lobby, you know, to look up laws yourself. I mean, it's such like an advantage on that end of things. So it's been great. Like so it's not smooth sailing, but it is like, you know, it's a big advantage for me to be able to understand this, you know, the legal side of things.
Ab (30:42)
Yeah, it makes a difference. Yeah.
Amazing. Yeah. And then I guess last question and we'll go into the quick fire round and start sort of wrapping up. you've done a lot of other things. So I think you've been able to just stack your skills over time to get to that place. I think a lot of attorneys and lot of people in general, you know, running a business, it's not the same thing as running a practice.
It's not that dissimilar, but certainly running a SaaS company, you're thinking about very different things to running a practice. Talk to me about that transition. How was that for you?
Mike Brown (31:20)
Yeah, well granted so both I'm running by myself at the moment and then I am, ⁓ you know, I know for cross beam I know we'll need to go hire people, but it is a truly interesting time period because how much work that these agents can do. So I'm kind of just pushing myself to see how far I can take it with just Claude, my open Claw, my open Claw's name is Laird. Yeah, he handles a lot of my admin work for me.
Ab (31:39)
Yeah. Yeah. Yeah.
Mike Brown (31:44)
So it's truly like a new age. And then I've been reading up a little bit more about company formations, those types of things. ⁓ Jack from the former founder of Twitter who runs Blocks nowadays, which is Cash App and those things, he wrote a really interesting concept of like this new age of if the AI has full visibility into, your meetings are on transcripts, your interactions in your company are on Slack.
Ab (31:51)
Yeah.
Hmm, yeah.
Mike Brown (32:12)
you know, every like every employee they're working with an AI. So you also have those traces. You know, the concept of like, essentially, you kind of get rid of middle management because middle management exists in a way to, to kind of just coordinate people. then it's like, all right, because they're already, you know, I wake up every morning and my, my Claude bot basically gives me my to do list based upon my emails and my GitHub, which is where my coding lives.
Ab (32:17)
Mm.
Mike Brown (32:39)
And just goes through and it's like, all right, yeah, here's what's on the docket today and calendar, obviously. And that's kind of like, I've been operating this way for about three months now and it's a lot better than the way I used to do it.
Ab (32:47)
Hmm. Yeah. Yeah.
Do you think do you think you lose a sense of agency just to go to into a slightly deeper question because now your day is being planned for you? How much? I don't want say flexibility because of course in this world you still have control for now. Yeah, do you think you lose that sort of sense of agency and that sort of spontaneity in your day?
Mike Brown (33:03)
Mm.
Do the AIs kind of control my day or? No, actually prioritize it allows me to do what I do best, which is it, you know, I've never I've always struggled with paperwork, repetitive tasks, more things that kind of bore me. I've always kind of, you know, I've gotten in trouble in school from having a little bit of a little ADHD personality, you know, bounce around. I like doing those things. And so with the with these AIs, I can basically offload all the things I don't like to do and focus in on the things that really
Ab (33:13)
Yeah, yeah.
Mike Brown (33:40)
you know, what make me wake up and be passionate about, know, being creative, figuring out the new thing, the new model drops, I get to benchmark things, you know, and then something like this, you know, and whereas, you know, I used to, you know, get a parking ticket and forget to pay it. It would be just be sitting in my brain and like, I need to do that. And I'm like, I'll do it tomorrow. So.
Ab (33:45)
Hmm.
Yeah.
Right, yeah. Yeah,
yeah. No, I agree. OK, yeah, perfect. All right, let's move into the quick fire round. I'm going to ask you a couple of questions if you're OK with it. They're quick questions, but your answers can be as detailed as you want them to be. What is one AI tool that you use every week that you'd miss if it disappeared immediately?
Mike Brown (34:21)
But Wisprflow We already talked about. mean, that is pinky to the function key. It makes my life so much better. Apple will probably end up buying them one day because it is so much better than Siri. But ⁓ until that day happens, please don't screw it up, Apple. But I love whisper flow. On top of that, Claude code is it just changed my life and not just because I won the hackathon. It just the day I started playing with it, it just felt like I something understood me in this world. And it was just like, all right, let's work together, bro.
Ab (34:27)
Yeah.
yeah.
Yeah.
Hmm.
Mike Brown (34:48)
Like it is so cool and good. And so I love cloud.
Ab (34:52)
Perfect. What one task in your job, your day to day job, I guess, or whatever you spend your time doing today, do you think will be fully automated within, I like to say two years, but let's say less even probably six to 12 months at the pace of change.
Mike Brown (35:08)
Yeah, I know I'm supposed to answer the question, but I'm going to flip the question because what task in my daily life should stay with me and not automated in a way is the way I think of it. so I honestly believe in two years, 95 % of the things I do will for the most part be automated. And that is the way I operate where I create skills. create cron jobs, are basically just like alerts to tell my AI do something at this time of day, right?
Ab (35:11)
That's it.
Mike Brown (35:35)
so I constantly am trying to automate everything, in a way now. So from like the legal perspective, right as a lawyer, cause I still do, uh, take cases on from the PI side. It's mostly referral work out there nowadays, but I genuinely enjoy being able to help people. know, someone gets in a car accident, they get hurt at work, something bad happens. They're in the hospital, you know, they can't make it to work. They don't know how they're going to pay their bills. If someone calls me, I will almost always drop everything.
to take that call and be able to help them for the next 20, 30 minutes and get them connected to the right people, right? And with AI, it's actually made me even better at that because like, so my company, Onbreeze is just a note taker, right? It fills out the forms that are necessary to go to different legal firms. And then through that, then my AI then knows, basically sees the notes and it just tells me who I should send it to. So anyways, it's made me a better lawyer in that sense because I could do what I'm actually.
Ab (36:12)
Hmm.
Mike Brown (36:28)
good at in my opinion, which is being able to be empathetic to people and help them and get them to where they need to go. So I will, no matter how good the voice, I know you love the voice agents, no matter how good the voice agent get, I'm still gonna answer that phone call. Because that's the one part I like about being a lawyer.
Ab (36:39)
Yeah.
Yeah.
Yeah. And, you know, I love voice agents, but I think there is definitely a differentiator for human empathy and for listening and for understanding and that connection, right. And there's a, yeah, there's a deeper philosophical argument of why people talk to AI so much talk or type because they feel heard, they feel seen, they feel understood as you were saying, you know, about Claude that it gets you.
And I think, you know, it certainly helps, especially when you're going through something that's stressful or a life-changing moment. So completely understand that. What is one of the biggest mistakes that you see teams, could be legal teams, could be businesses, making with AI at the moment?
Mike Brown (37:28)
Yeah, lock in, right? And I mean that as in everyone, do consult for a fair amount of different companies now. Everyone's trying to ask what is the best, what should they go get? And to me, it's like, look, what is best today will not be the answer next week. It is moving so fast if you're paying attention to this. So you need to be trying everything and you need to basically kind of whatever way you're getting your SaaS set up and stuff like that, you need to be open for everything and be constantly experimenting.
I would also say when it comes to token budgets, you everyone wants that budget number. tell people to spend as much as possible on tokens because it is such an unlock and you will see productivity go up. And at some point, yes, you will create a ceiling. You'll figure out where's a reasonable amount. And you don't want to create one of these environments you're hearing about in meta where people are writing Python scripts to burn tokens so then they could get raises. what I'm talking a lot from like you guys aren't even close to that point yet.
Ab (38:03)
Yeah.
Mike Brown (38:20)
We need to get people using this and you're gonna see quick productivity gains. yeah, just kind of, you know, try everything, be open for everything. Like just understand it's gonna 10X what you guys are doing every day. So, but you need to go through that learning period. So just get going with
Ab (38:20)
Right.
Yeah.
Perfect. And I guess related to that, what skill do you think legal professionals specifically should focus on building to stay relevant in the future?
Mike Brown (38:46)
Yeah, so like I said, with law school, as a lawyer, you are going to be naturally good at prompting, in my opinion. You spent three years of your life and spent a lot of money learning IRAC. And there is no place in the world where that's actually useful until the last year, where all of a sudden AI ⁓ agents actually respond really well to issue rule analysis conclusion when you give them the prompt. So.
You're gonna be naturally good at that. What I would start paying attention to is context management. The quick breakdown of that, I would look up this thing called Ripper 5, R-I-P-E-R, and then dash five, the five doesn't matter. It's research, innovate, plan, execute, and then review. Those five steps, if you can just put an agent through a loop like that, using that prompt engineering of ⁓ IRAC, you can kind of figure out that whole loop of things. You're gonna be,
at the top 95 % of people out there prompting and working with AI agents every given day.
Ab (39:45)
Awesome. And then last question, if you were starting from scratch today, where do you start?
Mike Brown (39:52)
Yeah, you just got to be curious, right? You just got to go do things. To me, it's ⁓ find something that you are passionate about, but there's low stakes. For me, that was fantasy football. And just go figure out a way to try to get you better. Use AI to get better at it. Ab, you brought it up earlier.
If you do, if you immediately are like, I got this case coming up and we need to do a bunch of research and you go have the AI do it, to be honest, you're lacking the skills to have the AI do well on that. So you need to find a way to work through something that's fun for you. And that will get you through kind of like the tougher parts of understanding context, ⁓ engineering, and then you go into the deeper parts. All of a sudden you're working with MCPs and connecting different APIs, and then you figure out skills, skills.md, capital S.
and then all of sudden you're automating things and you're just going to quickly you're going to get really good at this if you start with something fun.
Ab (40:49)
Perfect. Mike, thanks so much for coming on. I'll include all the resources in the show notes. If people want to reach out to you, if they want to find you, where can they go?
Mike Brown (40:58)
Yeah, so lately I've been all over LinkedIn. That's actually been fun. Since I won the hackathon, I've met some of the coolest people. And it's just for a year, it was kind of a lonely year of just being this lawyer who was just obsessed with AI. Turns out there's a lot of us out here, just like you, Ab. So it's just been a blast. So find me on LinkedIn. I'm Michael T. Brown. Ab will link to me, I imagine. ⁓
Ab (41:18)
Perfect. thanks so much for the conversation. So much fun and looking forward to seeing what you continue to build.
Mike Brown (41:23)
Absolutely, Ab Thanks so much for having me.