Runpoint: AI Business Transformation Podcast | Episode 2 | PowerPoint Is Dead: Building the Future of Work With AI

keywords
AI, tools, development, PowerPoint, automation, coding, data analysis, business insights, technology, productivity

summary

In this conversation, Matthew Hall and Sam Gaddis discuss their current projects, tools they find useful, and their opinions on the future of presentations and AI in business. They explore the effectiveness of various AI tools for automating tasks, the decline of traditional presentation software like PowerPoint, and the evolving role of AI in financial analysis and project management. The discussion also delves into their development stacks and best practices for coding and project planning.

takeaways

Sam is excited about a new project involving data scraping and analysis.
Matthew is focused on developing coding tools and models.
They discuss the utility of AI tools for business applications.
PowerPoint is becoming obsolete in favor of more dynamic tools.
AI can effectively replace some traditional roles in finance.
Customization of prompts is key to getting the best results from AI.
Planning is crucial before starting any coding project.
AI can significantly reduce the time spent on tedious tasks.
Using up-to-date API documentation is essential for successful coding.
The conversation emphasizes the importance of clear communication in project specifications.

Chapters

00:00 - Introduction and Exciting Projects

03:23 - Tools You Can Use Today

17:58 - Hot Takes on PowerPoint and AI

30:55 - In the Weeds: Development Stack and Challenges

Creators and Guests

Host

Sam Gaddis

Serial tech entrepreneur with a decade deploying practical AI. Experienced in sell-side M&A, he built and exited two companies: an AI products firm and a software consultancy where he served as CIO. Sam translates cutting-edge tech into practical tools aligned with business realities, advising and building buy-in with executives from Fortune 500s to startups.

What is Runpoint: AI Business Transformation Podcast?

Hosted by Runpoint Partners’ founders Sam Gaddis (tech entrepreneur & AI builder) and Matthew Hall (PE operator & growth strategist), Runpoint Podcast strips the hype from artificial intelligence and shows you how to turn it into concrete business results—fast.

Matthew Hall (00:00)
and welcome to the Run Point podcast. I'm your host, Matthew Hall, and this is Sam Gaddis. How you doing, Sam? Good. I'm doing good. We had a good week. I think things are going great for us. What do you think?

Sam Gaddis (00:06)
How's going? I'm great. How are you doing?

I'm loving it. This week we got a deal that I have been wanting to do for, honestly, since TragedyPD came out that involves scraping data from public hearings, transcribing the audio, analyzing that. don't know why. This is such a nerdy niche thing that I've wanted to do for so long and I finally found the right client for it. So I'm pumped to be able to work on that.

That's my thing that I'm excited about this week.

Matthew Hall (00:37)
I didn't know you were such a C-SPAN

guy. You're a public hearings kind of guy.

Sam Gaddis (00:41)
I just feel like there are these community meetings all across the country and it's mandatory that they record them and post them on the internet. And I feel like there's got to be a treasure trove of data in there for somebody and you just have to find the right person. I don't personally care about any of that data, but I know that somebody does. And I've just been wanting to mess with that for a long time. Like, feel like you could, I feel like you could programmatically scrape all of the city council meetings across the country.

Matthew Hall (01:00)
That's crazy.

Sam Gaddis (01:07)
and use that to understand what's happening in the country. Like where are there water quality issues? Where is crime popping off? And do it in a cadence that is completely outside or orthogonal to crime reports or the other bureaucratic institutional stats that you get.

Matthew Hall (01:22)
That sounds great and useful, but I think you could also do like just sentiment analysis and track like crazy hot maps of just where are things really popping off all around the whole country, right? Yeah, yeah.

Sam Gaddis (01:32)
Yeah, exactly.

So yeah, that's what I'm nerding out on this week. How about you?

Matthew Hall (01:37)
No, I've just been, we've been digging in, developing more stuff and kind of brushing up those chops. And it's always fun to see kind of who's on top in terms of the different coding tools out there and which models are working well. And it feels like your whole stack gets swapped every two months. So we're there right now, but we'll talk about that later. We've got three things on the agenda today, unless you want to add something. Our first segment.

we're calling, stuff you should use today. These are tools and tricks of stuff that's available for pretty much anybody out there to find utility with AI in their work right now. Second is a segment we're calling hot takes. Hot takes are just going to be our strongly held opinions that might be wrong, that are new. We're trying to be a little bit earlier, I think.

than the rest than the consensus on our opinions here. Wrong and wrong, yeah. And then the third we're calling in the weeds. So this is where I think the layman is welcome to leave the podcast would be what I would say, because we're going to talk a little bit nerdier about our own frustrations, what we found in building specifically with kind of vibe coding tools and model specifics. So those are our three segments. What do you think about that?

Sam Gaddis (02:27)
Strong and wrong, I like it.

Cool. Yeah, I've got a bunch of

stuff to talk about there that I'm excited about.

Matthew Hall (02:50)
Cool. All right.

Sam Gaddis (02:51)
All right, why don't

you kick us off with stuff you can use today.

Matthew Hall (02:54)
Alright, sounds good. Let me share my screen.

So generally, my thing you can use today is Lindy, which is what I talked about last week. But more specifically, I just want to demo a SDR agent that I built. a sales development. Is that what it stands for? Sales Development Resource? It's SDR, I mean.

Sam Gaddis (03:11)
Rep? I don't know. Sure. I've been saying that for 20 years but I have no idea what it stands for.

Matthew Hall (03:13)
Yeah, it's just one of those acronyms we all use.

Right, but it's outbound sales. Okay, and so the way that this agent works, and I'm gonna go ahead and start it working while I talk you through it. Every time a row is added to this Google sheet, it basically does research on the person and writes an email to it. And it's gonna draft the email and tell me to okay it before it sends to make sure it's not sending crazy stuff.

So I've actually put your brother's information in here and I'm going to copy into this row. And as soon as I did that, ⁓ it triggers this Lindy over here. And you can see in the task that it should be spinning up right now.

it is. Okay, so it's got it. So while it's working, I'll walk you through what the agent does. So as the new row is added to that sheet, it first searches its knowledge base. So you can add knowledge for the agent ahead of time. In this case, I have it look through our website, so runpoint.ai, so it knows what we do. Then it spins up an agent. The agent has this prompt where I'm telling it the world's best outbound emailer.

Sam Gaddis (03:55)
Check this.

Matthew Hall (04:17)
and it's acting on behalf of me, Matthew Hall, partner at RunPoint, and it's giving a lead that's a well-qualified executive in private equity. And what I want it to do is research the lead, use what you know about RunPoint and Matthew, and identify a great email hook. Write an email that is casual, punchy, and to the point. So this agent has exit criteria, so it can't leave until it feels like it's got a good amount of information and it has a good email. And so it's allowed to search the internet, it's allowed to pull from LinkedIn if it finds it.

and it's allowed to actually draft the email itself. So let's go back and check to see how far it's gotten.

disappeared.

All right. There it is. Hopefully we can fast forward through some of that, but so there it is. Right. You can see the agent working. So it's first search the knowledge base. looked up run point and myself, then search the internet. It found Ben's LinkedIn. You can see right there.

Sam Gaddis (04:58)
It's kind of cool to see it check along though.

Matthew Hall (05:10)
Now it's searching about the company itself. My problem with this is that it does too many searches. So I think that's probably something I can change in the Flow Editor, the number of things it does. But it seems to, you know, it burns through credits by continuing to search. And you can see if I hover over it, it gives estimates here about what each step takes in terms of credits. It underestimates though. I can tell you that once it's done.

Sam Gaddis (05:34)
I still maintain that at the scale of business that we are talking to, credits don't matter. A $500 a month expense. Like when you were maxing this out, was $300 or something, right?

Matthew Hall (05:47)
Right. So, okay, so here we go. We've got a summary here, which I'll ask you to confirm. Ben Gaddis is a partner and co-founder of SuperStep, previously served as CEO and president of T3. This is what they focus on. And now it's crafting an email. And it's got directions on this. Now, here's the draft of the email. And I've got it set to always ask for confirmation before sending. You can turn that off. So if you feel like you've got a great prompt and you've got trust here,

This thing will just send emails on your behalf and you can have it respond on your behalf and schedule the meeting on your behalf and do all those things. But here's what it said. It noticed a recent investment. Congrats. Your focus on next gen tech services caught my attention. Run point, we're helping PE firms deploy AI that delivers measurable results in portfolio companies, not in quarters, but in weeks. So it's taking stuff from our website. We recently built an automated tool. So it gives some sort of credibility there.

given your background scaling. I mean, what would you rate this? know, letter grade. I'm going to be completely honest. It's a B minus? Yeah. Yeah, yeah.

Sam Gaddis (06:44)
It's a B-.

That's it, you feel the same?

Matthew Hall (06:49)
I do. I think it's a B minus and I think with some work and crafting you could get it to a B and I think that's probably the top of it. know, but is that and so giving B grade ⁓ effort, you know, you know, whatever number of credits of dollars, that's what we're after right now.

Sam Gaddis (07:05)
I actually don't know. I think you might be able to get it to B plus or A minus because we haven't even messed with the prompt here at all. And my initial reaction to this is I don't know that the hook works. And it's also, you told it it's the world's greatest outbound guy or whatever, and it's writing like a very average outbound SDR. And so I think just modifying that prompt and giving it an example of what a good email looks like.

would go a really long way.

Matthew Hall (07:35)
probably true and you can pick the model I think 3.7 is generally the best at writing so that's why I chose that one but

Sam Gaddis (07:41)
Can you do four

or five? Can you do a...

for five.

Matthew Hall (07:43)
4.1.

4.5, yeah you can.

Sam Gaddis (07:44)
It's just very expensive. ⁓ I like four or five for writing. But anyway, I think you could get this to an A.

Matthew Hall (07:46)
Yes.

Well, that's the challenge.

Sam Gaddis (07:53)
We have to mess

with it. We should try that. We should run a campaign with this on those Dallas PE firms and just see what happens. So if you're in Dallas and you get an email from us, this is how it happened.

Matthew Hall (07:59)
Yeah, I agree.

I have fully adopted the Lindy meeting schedule assistant, which I just find more human than Calendly. It's just an agent that responds whenever I CC it and it can find times on my calendar or it can, it can just talk back and forth with the person to find the time. So it acts as like an EA as it goes to Calendly. Yeah. It's great.

Sam Gaddis (08:21)
man. I really want that.

I hate cal- Calendly is such a good functionality, but for some reason it feels rude to send that.

Matthew Hall (08:30)
I agree. You could rename it from Lindy too. So you could just give it a name of an EA, just pretend it's somebody else. And I don't think anybody would know.

Sam Gaddis (08:37)
I love that. Okay, cool. I'm going do that. All right, awesome. I actually really like that. I'm more impressed than I expected with that workflow. I have not seen these agents live up to the hype so far, especially the ones... I've seen agentic stuff that is custom built do really impressive things, but these off-the-shelf tools have not impressed me yet. That was cool.

Matthew Hall (08:58)
Good. Let's make it great. Your turn.

Sam Gaddis (09:00)
All right, mine is not super dissimilar from that. Let me share my screen over here.

Okay, so this is from a tool called leverage.app. And it's actually quite similar to what you're dealing with on Lindy. I just found that it worked pretty well. So if I do the same thing, we'll take my brother's company, my poor brother's, just the guinea pig for all of our demos. And we'll say, this is an intro call. Get to know you.

Matthew Hall (09:24)
Shut up Ben.

Sam Gaddis (09:29)
And we'll say that we're talking to my brother and Matthew.

It's going to do a pretty similar thing. It has the same thing on the back end. Let me pull that up.

Matthew Hall (09:36)
How does it never lie to you?

It has a knowledge base.

Sam Gaddis (09:47)
Okay, so functions pretty similar to Lindy. You've got a WYSIWYG flow here. You can specify the inputs to the form and then specify where you're going to scrape that external content from. You just pop in the client website that you got from the form. Now, this next step takes our company context and all I've done is copy paste everything from the website. You could scrape that, but since it's always going to be the same, I just dropped it in.

And then you're using, in this case, Cloud 3.7, but you can pick any of these as well. You've got a custom prompt here, and then it generates your output. And so what that looks like.

is something like this.

Sales call prep. It's got.

of the same stuff that you found from Lindy.

Matthew Hall (10:31)
So personal connections. Sam, let me ask you.

Sam Gaddis (10:32)
What I like about this is it does the value

alignment between what their website says and what our website says.

Matthew Hall (10:38)
how would it know personal connections?

Sam Gaddis (10:39)
It doesn't. We might be able to inject LinkedIn in there, but I didn't bother to do that.

Matthew Hall (10:47)
Gotcha.

Sam Gaddis (10:48)
Strategic talking points, pre-handling objections, gives you some questions and recommended next steps. How would you grade this?

Matthew Hall (10:57)
Well, I think it's all just about customizing the prompt. So I think it depends on what the individual seller or whatever you are, what do you want to be briefed on, and what are the sources of those things. I think if you can verbalize that succinctly, you can get whatever you want automatically from these things.

Sam Gaddis (11:15)
I agree. think these two tools are fairly similar and just applied in slightly different ways. It'll be interesting to see which one emerges as the easiest to use and that's really, I think what's gonna matter.

Matthew Hall (11:27)
Yeah, there's more every day. But yeah, I think there's value there. It's just who's gonna take it. All right, that concludes our Stuff You Should Use Today segment. We're gonna move on now to our hot takes. Hot takes of the week. Shout out Stephen A. Smith, the hot take god. Shout out the ringer. And so, my first hot take, I'll go first if you don't mind, is very simply that I think PowerPoint is dead.

And I don't know if it quite knows it yet. think what we found early on in this business is that it is easier and more effective to go straight from transcript or outline to fully functioning website. Fully functioning is loose term there, you know, but a front end of a website than it is to build any sort of presentation in PowerPoint or Keynote. And I mean, and so as a person who spent most of his career in corporate America,

It finally feels like it's my day in the sunshine. I've always hated PowerPoint. I've always been an outline guy. And some of my colleagues over time...

feel completely comfortable on a Blake PowerPoint screen of just like getting their ideas from here to the PowerPoint. I've never been able to do it. I always first have to outline diligently, you know, what the headline for everything and what I want to actually convey and then think about what the visual is and it's very linear in my mind. And so that is the perfect input for LLM and for coding. And so I just find it so easy and natural to go from that. So as an example of this, I'm going to share my screen again.

So as an example of this, have more or less, I mean, so far, we're doing our proposals ⁓ via websites. Websites that come from our outlines of what we want to propose and we give a structure to it and we pass it to a web development tool, in this case, lovable, but I don't think we'll be using lovable anymore since the latest update. But the concept is clear. I've made this anonymized so we're not sharing anything from anybody we're talking to. ⁓

It is we take the classic stuff in a proposal, what we heard, what the key questions are, what are the KPIs we're trying to improve, KPIs we're trying to improve. We bucket those things and this is kind of classic consulting stuff. We organize projects in a matrix of impact versus effort that's interactive so you can.

dive into each of the projects, you can see our understanding of them. You can, in a live environment, you can actually alter the values of these things, which change where they are on the matrix, ⁓ and move straight from here into recommendations for the proposal itself. So you kind of co-create the projects you're working on with a client. So it's very much the same motion as you would in a workshop or a PowerPoint presentation, but I no longer have to open up those softwares, and I couldn't be happier about it.

Sam Gaddis (14:10)
Yeah, that's so cool. I think for me, the most exciting thing about this is this looks like it took a lot of work to create, but we should be honest, it didn't. I mean, most of the work, all of the thought work happened in actual dialogue with the client or after the fact between you and me. And we're simply recording that and then having GPT take that transcript and make it into something that's organized and then taking that organized set of words and throwing it into cursor and saying, make this.

Matthew Hall (14:36)
Exactly, exactly. that's the, it is this, the hard work is still being done. The hard work is in the analysis, in the actual, what is the proposed recommendation for this client? What's it cost? What's the scope? What are we going to use? All that stuff. But that gets done.

in a conversation and then in an outline. And then from there, the actual visual interpretation of that and whatever kind of needs to be collaborated on can be done by the AI. And so it takes out the most laborious, least fun part of the process for me, which is always slide creation. Yep.

Sam Gaddis (15:11)
all the tedium. Yeah,

so let me show you one that I did as well. This is a similar thing, except in this case, we knew exactly what we were building. I've anonymized it in this case too.

Can you see this?

So here we had a call with the client, transcribe, we came up with what we were going to build for them on the call, transcribe that, threw it into cursor. There was some tweaking here for sure, but this just very quickly created exactly what we were agreeing to. And it even went so far as to create the ability for them to sign it and then download the executed agreement as a PDF and send it over to us.

Matthew Hall (15:39)
Alright.

Sam Gaddis (15:48)
And this took me all of maybe 10 minutes, and that's to create it from scratch. When we do this for the next client, it's going to take no time at all.

Matthew Hall (15:58)
Yeah. It's all it's this. So this is rather than the proposal creation, the like scope of services documentation, which again is just the least fun part of business. It's like, how do you distill what you talked about and we're having fun talking about with the client into some form of.

project management legalese. I've had writer's block a thousand times staring ⁓ at a word doc trying to figure out what are the assumptions and dependencies and how do I actually verbalize this project in a way that the lawyers are going to okay. And it's like that you can now get to 80 % of the way there and become editor as opposed to generator of these things.

Sam Gaddis (16:37)
That's awesome, I love it.

Matthew Hall (16:38)
All right, so it sounds like, yeah, no surprise here. We don't disagree on our hot takes. I would love to hear in the vibrant comments section if anybody disagrees with that take. The PowerPoint is, if it's not dead, it's dying.

Sam Gaddis (16:50)
I've got another one that I think might upset some people.

Matthew Hall (16:52)
Chew.

Sam Gaddis (16:53)
So I don't know if I talked about this last time, but I have a wood shop And this is like a commercial fabrication shop here in Austin. we don't have a finance team. And we're constantly trying to get these projects done for hotels and restaurants and things like that. And most of the people there are woodworkers.

Even the CEO is originally a woodworker. He's an engineer too, but he's a woodworker.

These guys are looking at QuickBooks, they're looking at P &Ls, they are doing their best to analyze this stuff. I'm doing my best to look at it as well. But these are businesses that none of us have ever run. I've never run a wood shop and the guys there also haven't. What we can do now that is blowing my mind is taking these exports directly as PDFs from QuickBooks and throwing them into something like Google's

2.5 experimental, which we talked about last week, and simply asking it questions that you would ask a CFO, and it is giving 100 % CFO level responses back. So for instance, I'm asking it right now, what is an appropriate gross margin for this particular type of business? How are we comparing over the last three months? What trends are we seeing? And then really importantly, given our target of a 15 % net margin,

What do we need in terms of revenue? And it takes into account what is our cost of goods sold for that? What is our variable cost? How have those things changed over time? How are they trending? And gives us not only what our expected revenue needs to be, but also like you need to keep your cogs at this level. It's just phenomenal. And I've back checked the math. Like it's real, it's actually working. It's doing exactly what you want.

And the way that this is happening is all through, it's all text-based. It'll write out the math for you, and so you can see it and check it, but it's not happening in a spreadsheet anymore. It's literally just happening in the chat. And I haven't seen models be able to do this until 2.5. I think 03 could probably do a pretty good job of it as well. But that's blowing my mind. And so I guess my hot take is, why would you ever have a CFO if you can just have a 25 or 30-year-old kid who knows how to use these models?

running your finance team. I mean, I know that sounds crazy, but in our resource constrained woodshop, I 100 % want to do that.

Matthew Hall (19:08)
Well, I think you could also inverse that. It's like, ever have junior accountants ever again just to have the senior person who's super-powered now because they can do all the grunt work as well? I think either way, you've got some long-term problems of people not rising from junior to senior. But that's a societal problem that might not be the topic of this. tell me a little bit more about, I got two questions for you on that. One.

You, I mean, you kind of downplayed, you you're not a CFO, you're not going to workshop, but you know your way around a spreadsheet. You know, know your way around business metrics. You know your way around those sorts of things. Do you think, what would it, if you put yourself in the shoes of, for instance, your colleague who has like been a woodworker most of their career and hasn't actually dealt with CFOs in the past, what would the gap be and how would they fill that? Do you think they'd still be, what would it take for them to get the same utility out of it you get?

Sam Gaddis (19:56)
I'll put it this way, one of the guys, when I was talking to him about it last week, pushback, he's one of these fellows who is a little bit skeptical of AI, totally understandable. And his concern was, it going to hallucinate, is it going to give us data that's inaccurate? Like I said, I back checked this. I really want the data to be correct. And so I looked at it. It was far superior to what I could do in a week.

If I dedicated an entire week to answering these pretty difficult questions of what should the expected revenue target be given these very specific constraints, that would have been a difficult problem. you're right, I do have some experience in business. I can answer that question. I know that I can answer that question. It answered it better than I could in about 10 seconds. So, yeah, I I just can't...

I can't quite wrap my head around this. And I am at the point where, and I've told this to our CEO, I'm at the point where I'm saying, look, you or somebody on the team has to become an expert in AI. Like it's absolutely mandatory. I know it sounds crazy because it's a wood shop with 10 people that work there, but it just seems irresponsible not to be using it.

Matthew Hall (21:06)
Yeah.

All right. Well, and then the second question, which has kind of probably made a good segue into the weed section. So can LLMs do math now? For a long time, that was the hang up. You had stuff like Wolfram Alpha that could always do math with natural language prompts, but you couldn't. GPT fell down constantly. Can the newest ones, you can trust it not with writing the formulas, but with doing the actual calculations, doing the math?

Sam Gaddis (21:30)
Yes, and with writing the formulas themselves. In fact, I've done a couple projects where I will have it create models. I was really interested in divorce rates and how those stratify in the US given different demographic parameters, so education level, age, age gap between couples, and all of that data exists in academic research.

But there is no model that you can go find that puts all of that together and says, given these parameters for two people, what is their likelihood of divorce? But absolutely the models can create that and it's creating that original math. And yeah, and it's good. I've yet to find 2.5 make a mathematical error. I'm sure that it does. I'm sure that it's possible, but probably not more than humans.

Matthew Hall (22:18)
Just the reasoning ones, so 2.5 and 3.0, can the other ones, can 4.0 can't do math, right?

Sam Gaddis (22:23)
For 0, I would not trust to do math. 0.3 and 2.5 is what I use for anything in that regard now. I have not put 0.3 to the test. For some reason, I just don't trust it quite as much.

Matthew Hall (22:35)
Yeah, yeah, same. Okay, cool. So like we mentioned before, that concludes our hot takes section. We're moved into the weed section and talk a little bit more about our trials and tribulations actually working with stuff. So first question I'm gonna put to you, Sam, is walk me through your current kind of development stack. What are you using right now to go to start a project, to spec it, to code it?

and what do you make sure you have in your setup to be successful?

Sam Gaddis (23:02)
Sure. Two categories. There are for fun projects that are just dead simple, that model thing. If you wanted to create a divorce probability calculator where you put in your age gap and all that stuff, that is, here's a mathematical formula. These are the fields that we want to include. Make that. I'll throw that into Replet Agent and it'll pretty much one-shot that.

And then the nice thing about Replet Agent is you can deploy it from there. I built another thing that was like a guitar ear trainer thing where it actually listened to you play a note. So it would say, play E4, or play E on the fourth string. And from the moment that it said that, start timing you, and then it would listen for you to play the correct note and then give you the time and calculate those.

Those kinds of things that are well-defined and are pretty much just operating entirely in the browser, Replet Agent is great for. And you can go a long way with that. However, if you need some sort of database, if you need to build a production app, pretty much everything that we do for clients for sure is happening in Cursor. And my stack right now, there's a...

I would say 80 % of the effort goes into the planning upfront. And once you actually get to the coding or the vibe coding, you're really just pushing buttons. It's like kind of like flying a plane on autopilot and you're troubleshooting errors and stuff like that. the hard part and the part that people don't do enough is developing the spec, which obviously starts with what are we trying to accomplish? What are the goals? All that good stuff, all the consulting stuff. So...

It's lot of that, it's several iterations of that, starting at the highest level, what are we trying to accomplish, then going to a product requirements document, and then going all the way down to what are the individual screens and what are on those screens. And from there, what I will typically do is design each of those screens, and I've gone through a million different iterations of how to do that. V0 is a tool that I was using for a while. I know that you were using Loveable for stuff like this. I have tried the various...

know, interface design tools that are like Figma, but it's doing it automatically. I have not found any of those to be good. Right now, my current state of the art is finding an interface that I like, taking a screenshot of it, dropping it directly into Cursor, and saying, I want a screen that does these, here's the context of the app and what we're trying to accomplish, and make it look kind of like this. And it does a great job.

Matthew Hall (25:21)
And are you

using, what models are you using? like Replet, think 3.3, Sonnet 3.7 is its default, I think, but in cursor, you using Gemini or are you switching it up?

Sam Gaddis (25:31)
I'm using primarily 2.5 experimental in, or not experimental anymore, 2.5 Pro in cursor. And that's the reasoning model. It's fast, it's good. I like it. Context window is huge. The other thing...

Matthew Hall (25:43)
Although I've read like

it actually has a shorter context window than Gemini does. Like cursor doesn't have a big enough context. It's not letting you use the whole thing for token usage.

Sam Gaddis (25:50)
It's not letting

you know. Absolutely, it's not. But still, I don't find it to be limiting. The other things that I would say are, once you start getting into designing, there are a couple MCP servers that I found to be really useful. One is Context 7, which essentially gives you up-to-date API documents for anything that you're going to be using, including the newest models from OpenAI or Google or whoever.

But importantly, it also gives you code samples to go along with that. Another one that I'm really liking, let me just pull up my cursor here and look at it. Another one that I'm really appreciating right now.

Matthew Hall (26:24)
Before you can

go on to that, Sam, I want to underscore how important and cool that context seven thing is, just because anybody who's tried to vibe code in the past has invariably come up against a thing where...

The LLM is using outdated docs, right? And so it's, using docs from GPT from, from 2021, whatever model you have has its own knowledge cutoff, right? And so it thinks that it has the state of the art, but it's actually doing things that aren't just old or sometimes deprecated, right? And so it'll, it'll break your entire code base frequently, you know, and it can become incredibly frustrating. And so you can, you can, you know, give it more docs. can up, you can actually download the docs. can link to them or any of those things, but it also

seems to like not always pay attention to those and it all like it increased your context window all that kind of stuff so the MN MCP server that gives it the most acidic API docs is like it reduces headaches considerably so

Sam Gaddis (27:16)
Yeah, you spend way too much time on that stuff. And it throws you off course is the problem. I feel like the two keys to vibe coding are one, doing as much planning up front as you can, creating all of that documentation up front, and then running it off a plan. So you should have a list of tasks that it's running off of and a variety of ways to do that. But then the other thing is just getting ahead of all those troubleshooting...

those roadblocks that you run into, like the API mismanagement of old API docs and that kind of stuff. If you can prevent that, the other thing that would be similar to that is troubleshooting database errors because you're not doing migrations or you're not using the super base MCP or something like that. those things can eat entire days if you're not careful.

The name of the game is just to avoid those problems. Don't lose out on that. The other thing that I would add that I'm enjoying is OpenAI. It's called OpenAI GPT Image MCP. And this is so silly and simple, but it allows you to generate images using the new OpenAI 4.0 Image Gen and then drop them directly into your project, into your site that you're building.

directly from cursor. there's no more like, you can do everything within cursor. You can say, this website needs an image here and just tell it, make me an image that looks roughly like this and it pops it in.

I did it for the proposal. I said, give me a header that looks like abstract techy stuff. And there it is. The final one is one that I cannot completely vouch for yet, but I think looks so promising and it's called Taskmaster AI. This is an MCP that does...

a really robust version of what I was saying a second ago, where it outlines all the tasks and you give it your initial documents, ideally one big markdown file that says, here's what I'm trying to build. And it will carve that up into as many discrete tasks as it needs. Then it will rank those tasks in terms of complexity. Then it will, if a task is too complex, break it up into subtasks and you can chug through those one by one.

and as it's going, will continually update itself on where it is in the process and it doesn't get lost.

Matthew Hall (29:21)
Yeah, so I used that yesterday and it's interesting because it's...

That's basically been my workflow. And I think for a lot of people, that's kind of best practice in, in vibe coding is to kind of start with the law, just as you described a long conversation to get to the point where you get a PRD with, with an LLM and with your LLM of choice. And I think like the best way of doing that is you give it the format you're looking forward to PRD and then you have it ask you questions, clarifying questions until it has a good understanding of the project that you're, that you're doing.

I've fallen down a couple times in not spending enough time doing that or trusting that the doc was good based on the conversation without proofreading what the AI wrote in the PRD. And then later on down the project, I realized that like encoding it, it did what it was supposed to build, but it wasn't what I wanted because I didn't actually clarify what I wanted upfront. So it's like crucial. the taskmaster, the other, the next step of that is after you the PRD is to break down in your cursor rules.

specific rules that basically do what Taskmaster does. And so I played around with Taskmaster MCP yesterday. It basically just pre-bakes your cursor rules for you. It gives you like three or four files of, know, they're basically pre-prompt that the cursor LLM reads before it reads what you're asking it in the conversation, right? And it has a very specific format for its task list, which is very similar to the one I've been using, you know, anyway. like, and it always...

reads through them, marks off the ones done, moves pending to close, moves on to the next one. And it always importantly outlines what it's going to do before it does it, which is a pain everybody's had with cursor for a long time, is that it always jumps to changing your code when there could be a simpler way of doing this or a root cause of the bug it's facing or anything like that. So it's pretty cool. Yeah.

Sam Gaddis (31:03)
But it's still new. You gotta wrestle it into submission.

Matthew Hall (31:07)
I just don't, it's not that you can just have, and this is what I, you I have a bunch of prompts I save in a notes file, right? That like when I start up a new project, I copy and paste them to a new cursor rules. And this is the format of these things. And I do the same thing for starting a project. And there's a lot of these on the internet of just like, you know, I've just customized mine. It's just sort of those out of the box. And so it's free, it's a good reason, it should be free. It's not something I would pay for, but yeah, it's useful.

Sam Gaddis (31:33)
Cool, what else, is that it?

Matthew Hall (31:34)
I think that's it.

Sam Gaddis (31:35)
All right, this was great. I think we covered a lot.

Matthew Hall (31:38)
I think so too. Excited to see the clips. Goodbye everybody. Have a good weekend. See ya.

Sam Gaddis (31:40)
All right.

See ya.