How a senior staff engineer at Clerk actually uses AI to code

0:00Today I'm joined by a friend that I got

0:03to know a little while ago. Uh I've been

0:06collaborating with Clerk at a hackathon

0:09for Nex.js Conf. Brandon Romano is an

0:14engineer at Clerk. I have gotten

0:16introduced to him and his work um while

0:19collaborating for educational content

0:22but also got to know him as a person as

0:26a friend at the hackathon. we got into

0:30taking so much photos of the event. Uh

0:33maybe at one point I would also share

0:35that with the folks. Um seems like we

0:38have a kind of similar interest in

0:41photography, event photography or

0:43portrait photography. Brandon, uh I'm so

0:46glad that you're joined. It's an honor

0:48to have you here and thank you for for

0:50doing this.

0:51>> Yeah, thank thanks for having me. This

0:52is I'm like very excited to be certainly

0:54talking about this topic cuz I'm like

0:56very excited about AI right now. I think

0:58the industry is really really exciting.

1:00Um and yeah, also yeah, we we definitely

1:02discovered very quickly that we were

1:04kindered spirits for sure. Um you should

1:06totally put up some of the photos that

1:07we that we uh that we took on the

1:09screen. Um but yeah, no, that was that

1:11was definitely a fun it was a fun way to

1:13>> That was a fun night. That was a very

1:15fun night and um you know um we met a

1:18lot of cool people that were were

1:21attending at the hackathon building cool

1:23stuff

1:23>> and you could see the theme was AI. I

1:27was either an AI native or they just

1:29built the whole thing with the AI. So I

1:32thought to have you on primarily looking

1:34through the lens of somebody that's just

1:37working on the back end of an

1:39authentication company or technology,

1:42but I wanted to know how you as a senior

1:44engineer

1:46um use AI. What's your day-to-day

1:48pipeline? What's your workflow? What

1:50tools do you use? How do you harness it?

1:52How do you see the advancement over the

1:53past few months? Where did you start?

1:56And how much are you using it now? And

1:58where do you seeing it going? So um

2:00let's just start with like your journey

2:02of using AI and then maybe at the end um

2:05I know you have also prepared a project

2:07so we can walk through an actual project

2:09build it from the back end and also go

2:11to the front end and show it end to end

2:13so we can kind of learn more about how

2:15you use it dayto-day.

2:17>> Yeah totally. So yeah, so so definitely

2:19from like my journey perspective, I

2:21think it was probably the about like

2:24close to the second half of 2024 is when

2:26I kind of really picked up, you know, AI

2:28tools in seriousness. Um I when GitHub

2:32Copilot originally came out, I gave it a

2:34shot and I was actually very like I was

2:37not happy with it. It felt like the the

2:39way I described it initially was like it

2:41was like a junior developer was just

2:42like screaming over my shoulder with

2:44like the bad ideas constantly. Um, and

2:47it was very distracting. Um, but you

2:50know, it's really evolved over time.

2:51Like I think probably at certainly at

2:53the start of 2025, the models were

2:55getting like really good and the

2:56harnesses were getting really really

2:57good and it and and I gave it another

3:00shot and I was like very impressed and

3:03kind of ever since then I've kind of

3:04been like slowly incorporating it more

3:06and more into into my workflow. Um there

3:10are you know days now where I will not

3:13type a single line of code and I and I

3:15delegate almost entirely to AI. Um it's

3:18quite good at working certainly in

3:20existing code bases. Um you know I do

3:22have to spend a lot of effort kind of

3:24like giving it particular directions and

3:26saying okay go and do this in this

3:28particular way and you know it goes in

3:30the wrong direction you need to correct

3:32it. Um, it's a thing that, you know, you

3:34still actively have to participate in

3:35and you still have to have the

3:37underlying fundamental knowledge of what

3:39it's actually doing to operate it in its

3:42most effective way. Um, but yeah, it's

3:44it's like profoundly changed um software

3:47development in in certainly in the tech

3:49industry. I know that, you know, we

3:51we're probably in a little bit of a

3:52bubble in the tech industry. We're very

3:54much like at the forefront of of

3:55adopting these types of tools. Um but uh

3:59it's certainly in the tech industry,

4:01it's every single developer that I'm

4:02talking to is very significantly using

4:04AI in their workflow and for very very

4:07large tasks. It's no longer just like

4:09write this function. It's like I'm

4:11building this feature like let's like

4:13come up with a plan together, you know,

4:15with the AI tool and and go and build it

4:17and you know can be pretty significant.

4:19Um so so yeah, so that's been my

4:21journey. It's like it's definitely been

4:23a story of using AI in very specific you

4:27know ways like kind of like a scalpel.

4:29Um and now it's kind of become a bit of

4:31like a you know more of like a

4:33sledgehammer that I can swing around and

4:35obviously you know with that with great

4:36power comes great responsibility. Um but

4:39it it really can get a lot done.

4:41>> Yeah absolutely. I feel like um it's

4:44very similar to my journey or and a lot

4:46of people that I talk to it starts from

4:49I don't know tab completion with GitHub

4:51copilot that's just like that and then

4:54maybe testing and trying agents here and

4:56there for a function for a little small

4:58change but then um I think what changed

5:01it for me was the plan mode in cursor if

5:03you're using um you can we it just goes

5:07back and forth and it asks questions

5:09before implementing and you could see

5:10how it's thinking or what like loop

5:12loopholes um and and holes in the plan

5:14there is and then it goes unexecuted. I

5:18think um specifically with the Opus 4.5

5:22um whomever I'm talking to they're like

5:24yeah we were using it. Yes, it it was

5:27before that that we you know we have

5:28rules, we have guidelines, we have

5:30agent.mmd, we have there's like we have

5:32harnesses that to just make it produce

5:34good work, but seems like everybody with

5:37the opus 4.5 and cloud code or open

5:40code, they're just like trusting it more

5:42and more. Um,

5:44>> yeah,

5:44>> but I feel like it's it's it's it's a

5:46very good model. It's definitely that's

5:48when I and I think a lot of my

5:49co-workers were were like, "Oh

5:53like things have kind of changed." um a

5:55little bit. Yeah. Sorry, I cut you off.

5:57>> No, no, no, no, no. Yeah, exactly. Yeah.

5:59I mean, I was talking to Tanner uh

6:02Linsley um few weeks ago and he was like

6:05the same thing. He was like um he was

6:08telling me that he was talking to his

6:09team and engineers about kind of the

6:12same experience that them being

6:14skeptical in the first not trying it or

6:16as you mentioned giving it a very short

6:18leash, not letting it just go um on its

6:21own. But then the 4.5 Opus 4.5 came out

6:25and then he was like I told them guys

6:27just give it a try and see and it just

6:30blew up their mind as well. Um but I

6:33think it's also even even if the model

6:34is good I think it is also very

6:37important how you're providing context

6:39to it. How what are your rules? What are

6:41your MCPs? Are you providing any skills?

6:43Are you providing any guidelines any

6:45best practices there? the the more you

6:49have these guard rails in place, the

6:51better the way better the quality of

6:53what you're getting. Uh and often times

6:56when I see people not getting good

6:58enough results, I'm like, "All right,

6:59maybe you're not providing good enough

7:02context or rules or skills or what is

7:04your what is it? What is in your agent

7:07MD file?" Um sometimes it's just

7:09accessing old documentation and whatnot,

7:11but there there are ways that you can

7:13make that better. And I think this is

7:14the new area of learning what everybody

7:17else is doing like yourself or other

7:19engineers in the industry and try to

7:21replicate and kind of establish best

7:23practices on how to use this.

7:25>> Yeah, absolutely. Yeah. I mean, so yeah,

7:27context is very much key. It's like it

7:29is only going to be as good as the

7:31information that you give it, right?

7:32It's really good at, you know, taking a

7:35bunch of information and processing it

7:36and coming up with, you know, an a

7:39reasonably good answer to whatever

7:41question you you propose it. But yeah,

7:42if you don't give it any context, if you

7:44don't give it those guard rails, um it's

7:47likely not going to do the thing that

7:48you want it to do. Because with with all

7:50software engineering, there's a million

7:51ways to do things, right? So which one

7:54of those ways I if you don't tell it

7:57which one of those ways to do, you know,

7:59the thing that you want it to do, you're

8:00just going to get one of those ways. And

8:02you know, maybe you might disagree with

8:03it. And if you disagree with it, you're

8:05going to have a bad experience with AI.

8:06Um so yeah, really it's definitely about

8:08the context that you provide it. Um, and

8:10you know the tooling is getting really

8:11good like as you mentioned there's you

8:13know there's there's all these you know

8:14context markdown files that you can have

8:16inside of your repo but also too just

8:18like even just the editors I think today

8:21make it super easy to provide that

8:23context like very often when I'm giving

8:25a prompt what I'll do is I'll say you

8:28know you need to do this particular

8:29thing and then I'll you know I think

8:31it's a command I don't even it's muscle

8:33memory so I don't even know that the

8:34memory is like command L or something

8:35like that where I could say like you

8:37know this reference this part of the

8:39code um and do this and I can give it

8:41very very specific directions and I can

8:42kind of like embed the context of the

8:44very disperate parts of my codebase um

8:47and kind of give it how I want it to do

8:49it and it it it is is very much um it

8:52definitely improves a the output when

8:54when you give it that right level of

8:56context.

8:57>> Yeah, I mean when you're dealing with an

8:59existing code base um what you just

9:02mentioned is is key. You're pointing

9:04you're like okay I want to build this

9:05feature and this is how I write my

9:07functions. this is how I write my

9:08endpoints. This is how I I don't know do

9:10this a specific thing. Go look at this

9:12and then implement it here which is I

9:14guess very good at just doing that

9:17versus for a new project where you're

9:19just like um there is no code reference

9:22or the way that you prefer you you just

9:24have to provide more context or rules

9:26and guidelines and guard rails for it to

9:27do it. Um, that's actually very

9:30interesting because I was talking to

9:31somebody yesterday with a a a CTO of a

9:35software company up here in Toronto,

9:37Canada, and they're like our code base

9:38is like about a million lines of code

9:40and it's just like um I guess C++ or

9:44something. It's like a ERP management

9:48like like logistics and you know

9:50inventory management kind of software.

9:51and he was like um we haven't had we we

9:55haven't been comfortable or confident

9:57that it understands our code base good

10:00enough. Uh and I also mentioned the same

10:03things that you said that you could

10:04point it at different you know um

10:06guidelines and you could create these

10:08different you know best practices or

10:09ways you're just like working or the

10:11architecture of your um you know

10:14software and then it just continues to

10:15produce that within that same system.

10:18[snorts]

10:19>> Yeah. So, so this is definitely a thing

10:21that, so if I were speaking to this

10:22person, the advice that I would tell

10:23them is that you can give the context of

10:27your architecture to, you know, the AI

10:30agent. You can and you can write a

10:31document yourself, human written. Um,

10:33and actually I have one such doc

10:35document in the uh in the codebase that

10:38I'll that I'll pull up later. Um, but

10:40basically like you know, all software

10:41projects, yes, they're complicated when

10:43you look at them from a whole. Of

10:44course, it's going to be a million lines

10:45of code. Um but you every you know every

10:49codebase you can look at it at different

10:50resolutions right you have the highle

10:52architecture and you say oh okay there's

10:54we have these [snorts] you know we have

10:56the HTTP layer and then we have a

10:58service layer and then we have a

10:59repository layer right so it generally

11:01can be kind of described generically and

11:03like here are the way that things are

11:04done here um I would say that uh these

11:08agents are actually they're quite bad at

11:10coming to those conclusions I would say

11:12generally the big picture stuff it's

11:14hard for it to get the big picture stuff

11:16because it it isn't going to crunch a

11:18million lines of code every single time

11:19you send it a prompt unless you want to

11:21use like a hundred million tokens every

11:22single time you know you send a prompt

11:24in. So you need to be able to guide it

11:26in such a way where you can give it that

11:28highle contextual information about like

11:30where are the what are the steps what's

11:33what are the layers in my application

11:34that I need to you know to look at

11:36whenever I make a change. Um and that

11:39generally will help and you can provide

11:40that in kind of like exactly like an

11:42agents.mmd file or you can just have

11:44like an architecture.mmd file and

11:45whenever you're performing a change

11:47where you're like I think it needs this

11:48you can link it to that inside of the

11:50prompt and provide that context.

11:53>> Yeah. Yeah, that makes sense. All right.

11:57Um,

11:59do you want to continue by actually

12:02diving in and walking as you're

12:04explaining like how you're setting up

12:06this project and you know different

12:08files, guidelines and whatever and I

12:10will just jump in, interrupt you if I

12:11had any questions otherwise I'll just

12:12let you uh take over.

12:15>> Yeah. Yeah, I think that that sounds

12:16good. Let me share my screen here. Um so

12:19yes so I have created a um let me share

12:23my screen first uh allow to share the

12:26entire screen share the entire screen

12:30okay cool so yeah so as I had mentioned

12:33um I've created a just like a fake

12:36project um just so we can kind of like

12:38go through and I want to you know show

12:41some real concrete examples of of going

12:43and running some prompts myself because

12:46I think seeing is believing with these

12:48types of things. Um, so the conflated

12:51example that I have put together is, uh,

12:53I'm going to be opening up a cat cafe

12:55with my cat. My cat's name is Luna. Um,

12:58and I need a system to, uh, manage

13:01tracking cats, tracking people that see

13:03them. There's going to be some adopt

13:05adoption applications. That's the goal

13:07of this cat cafe that we get cats

13:09adopted. Um, so you know to manage all

13:12of this I'm going to create a system uh

13:14to to manage the cats, the customers,

13:16the reservations and the adoption

13:18applications and then hopefully you know

13:20they're approved and then they become

13:21owners. Um, so I'm going to so in this

13:25I'll build a a backend and that's going

13:27to be the context. So you can let me

13:29know if I if I veer off in a particular

13:31direction and and you have any questions

13:33about you know particular things because

13:35I I very much have the curse of

13:36knowledge on uh on these types of

13:38things. So, like I will probably glaze

13:41over something that's really complicated

13:43underneath of the hood. And so, feel

13:45free to ask any questions at any point.

13:47Um, but so, yeah. So, I I've set up I've

13:50set up a project uh initially with with

13:53some boilerplate and a little bit of

13:55work I put into this um just because it

13:57would be extremely boring to to set up a

14:00project initially. Um but so

14:02specifically I I will say though about

14:04the project setting up process is so if

14:07you are working in a codebase AI works

14:10better in an existing codebase because

14:12it can infer what your opinions are

14:14about how things are structured and it

14:16doesn't have to make as many decisions

14:19right so in the process of setting up

14:21your codebase you actually want to how I

14:24mentioned there's kind of like the

14:25scalpel and then the sledgehammer you

14:27want to actually when you're setting up

14:28your codebase you do want to Think of it

14:31like it's a it's a scalpel operation

14:33because you're making pretty

14:35consequential architectural decisions

14:37that the uh the any future AI prompts

14:40will assume are codified inside of your

14:42codebase. So if I make some bad

14:44decisions up front about how this thing

14:46is structured. Um it it's going to be

14:49it's going to be poorly structured

14:51forever. So when you're setting up the

14:53project, you actually want to kind of

14:55like you know apply your knowledge that

14:57you have you have uh already gained as a

15:00developer. And if you have to actually

15:01go in and type some code yourself, it's

15:03not a bad thing. Um, AI really gets its

15:06power once it's kind of working within

15:07an existing codebase. Like you can

15:09totally spin up a new codebase and it

15:11can make reasonable decisions, but just

15:12make sure you review them and you agree

15:14that this is the right architecture and

15:16setup for your project. Um, so, okay.

15:18So, I'm going to I'm going to walk

15:20through very quickly some of the

15:22boilerplate uh and and the stuff that I

15:24have set up. Um, so as I mentioned, I'm

15:26going to build an API that's going to

15:27manage this. Um later on we're going to

15:29>> Brandon.

15:30>> Yeah.

15:30>> Brandon, can you make the screen a tad

15:32bit bigger? Like zoom in a little bit so

15:34that Yes.

15:35>> Yes. Okay. Perfect. Cool. Okay.

15:38>> Beautiful.

15:40>> Yeah. Yeah. Okay. Cool. Um so, okay. So,

15:45I'm going to jump through and just kind

15:46of walk through some of the boilerplate

15:49here. Um so, uh as I mentioned, so I'm

15:52building an API. Um, so this API is

15:55really really simple in terms of like

15:56its overall architecture. I have a it's

15:59just a Postgress database uh and uh and

16:02it's going to be a Go application that's

16:04going to be the uh the the code that's

16:06going to that's going to run it. Um so

16:08in it I've set up some very simple

16:10docker setup so I can go and spin up the

16:13application really easily. Um and as

16:16well I have like a source directory and

16:18I have stubbed out uh the the server. So

16:23I have got this server abstraction where

16:25I can go and say here are my routes and

16:28I can plug in those routes into my

16:30server and then go and build the

16:32services that go and contact the uh

16:35repositories which go and fetch the data

16:37from the database. Um so that's from a

16:39high level that's kind of the setup of

16:41this uh of this project. Um, so what I

16:45figured would be useful, uh, what would

16:48be helpful is I can kind of just like

16:49walk through some I could walk through

16:52building some, uh, some features inside

16:54of this and I can I'm going to start

16:56kind of small and then I'm going to kind

16:57of grow to giving it more and more

16:59responsibility.

17:00>> Sounds good.

17:01>> Um, so I think the first place that I'm

17:03going to go, so basically, so I've

17:05already built this endpoint to to create

17:07a new cat. Uh, and I've already built an

17:10endpoint to go and fetch a cat. And I

17:11did this because I wanted to give I

17:14wanted to be involved in this process

17:16and and give the codebase an example of

17:18like here's how things are are ought to

17:21be done inside of this codebase. Um so

17:23now AI can reference these other things

17:25when it's going and building uh new

17:27endpoints. Uh and it can use this as

17:30context. Um so I'm going to start really

17:32really small. So, I want to build the

17:34endpoint where I can go and fetch all of

17:36the cats um that are available for

17:39adoption or just all of the cats in

17:41general. Um so, the first thing that I'm

17:44going to have to do is um I want to

17:48let's see. I'm going to go and into my

17:51cat repository file. Let me move this

17:54out of the way. Uh I'm going to go into

17:55my cat repository. So the repository is

17:58the layer in the in my application where

18:00I am contacting the database and I'm

18:02going and say this is where I'll write

18:03my SQL queries and where I'll go and say

18:06give me all of the cats right so if I'm

18:11you know not so trusting of uh AI quite

18:15yet um I might only want to give just a

18:18single function level of responsibility

18:21right so I might want to just go and say

18:24uh you know I might want to stub out the

18:27actual function signature right so

18:28function uh r cats repository not create

18:32color I want this to be uh let's call

18:35this fetch cats right and this should

18:40return an array of

18:45models cat or an air

18:50right so I'm still very much human

18:53involvement so what I can do here And

18:56also too, I'm using cursor. I recommend

18:58cursor as a uh initial tool if you're

19:00getting started with AI. It's very it's

19:02a very similar experience to how you

19:05maybe used to write code prior to prior

19:08to using AI. Um there's obviously a a

19:11million different tools that you can

19:12use. Cursor is is quite good for this.

19:14So in cursor um what I can go and do is

19:16I can just go and say I can you know

19:18press what am what what did I press?

19:20It's muscle memory. So it's I think it's

19:22command L. Yeah, command L. And then

19:24what I do is I provide the line by line

19:27context inside of my chat. So and I can

19:29go and basically go and say implement

19:31this function.

19:33Uh I want to be able to fetch all of my

19:38cats from my database. Uh I have a spec

19:43I have a create cats table. So I have a

19:47uh a specific uh thing inside of the

19:50cats table uh called an owner ID which

19:53is basically this is like this cat has

19:54been adopted if they have an owner. Um

19:56so I want to be able to uh fetch cats as

20:01well that are available adoption. So, I

20:04want to uh be able to fetch cats who are

20:08available

20:11available for adoption,

20:14which is determined by the fact that

20:18they don't have an owner.

20:21Uh

20:23yeah, let's get let's go. Let's see what

20:25that goes and does. So, I've provided it

20:28the signature. I'm not giving it a lot

20:30of responsibility. I'm just going and

20:31saying build this repository function.

20:33Um and we'll see how it does. Um so I

20:37think right now you probably see that

20:38cursor is is pulling in some context.

20:41It's like okay let's go and understand

20:42what the cats model is like. Uh I think

20:45it reached into the create cats table to

20:48to see what the schema is to make sure

20:50that it can run that query correctly. Um

20:52so it's kind of like right now it's in

20:54its context gathering phase to be able

20:56to execute you know this type of query.

21:00Um, and so now we are building this

21:03fetch cats query. So we're gonna let it

21:05do its thing and we're gonna we'll

21:07review it.

21:09Okay, let us see how it did. Okay, so

21:13fetch cats. So okay, so there's there's

21:15a few options. Um, yep, it has the

21:17available for adoption option which I

21:19can pass in optionally filters which

21:21cats are returned.

21:24Pass nil for ops to fetch all cats or

21:26set available for adoption to filter

21:27availability. true, no owner, false, has

21:29owner. Um, cool. And it allows you to

21:31also to not fetch that. So, here is my

21:33base query.

21:36It goes and says if it's available for

21:37adoption. So, it's doing what I want it

21:38to do here. And it's going and querying

21:40that. It's scanning the rows of the

21:43database into the actual cap model. And

21:45it's returning the caps. So, cool. So,

21:47this is this did what I wanted it to do.

21:49Um, I didn't give it a lot of

21:51responsibility, right? So, if we look

21:52inside of our, you know, what did it

21:55actually go and do? uh it just modified

21:57this single file and it just did this.

22:00So I didn't have to I don't have to give

22:03AI a lot of trust because it's very easy

22:05for me to go and validate that this is

22:08the thing that I wanted it to do. So

22:12generally if you're getting started with

22:13AI or you're not using you know AI

22:15workflows currently um this is the the

22:17where I would say start start with a

22:21very small amount of responsibility

22:23because it's easy to verify. It allows

22:25you to like build up your trust that AI

22:28is kind of pretty good and can do the

22:30thing that you want it to do. Um and it

22:33kind of allows you to still remain in

22:35like really really tight control over

22:37everything. So this is kind of the first

22:40piece of the demo is just building this

22:41one particular function. So I'm going to

22:43commit this um get in

22:50and fetch cats

22:52function.

22:55All right, we're just pushing the main

22:57today um for now. Um okay, cool. So

23:02that's kind of the start of it. Um, so

23:05but what I'm actually after is I'm

23:07actually trying to fully build this

23:09endpoint. Um, so I've gained a little

23:12bit of trust that the AI did the thing

23:14that I wanted it to do. Uh, so I'm going

23:18to now just tell it to build this

23:20endpoint.

23:24All right. So I'm just going to say uh

23:27uh build this endpoint.

23:30Use

23:31the fetch cats

23:37use the fetchcast repository function

23:40uh to perform the underlying query. I'm

23:45going to also to I'm going to give it a

23:46little bit more context because I know

23:48I've used AI enough to know that it's

23:49going to try to implement pageionation

23:51because uh this is generally uh how how

23:55it goes and does it because that's best

23:57p practice. We're not worrying about

23:58best practice today. So, I'm going to

23:59say uh don't uh worry about

24:02pageionation. You can just return all of

24:06the cats

24:08that match for now. Okay. So, now we're

24:12giving it a little bit more

24:13responsibility

24:14and we are just telling it to build the

24:16entire function. And this is going to

24:18require a little bit. It's going to have

24:19to build inside of my inside of my uh

24:23application. uh I have so it's the HTTP

24:26layer and then there's a service layer

24:27and then there's a repository layer that

24:29goes and sends a request to the

24:30database. So it's going to have to go

24:32kind of up and down the stack quite a

24:33bit to be able to successfully build

24:35this endpoint. [snorts] Um it's going to

24:38have to register the the the endpoint to

24:40the uh to the router so it can actually

24:43go and handle it. Um treating available

24:46option or filter otherwise returning all

24:48cats. Cool. It seems to be doing what we

24:51want it to be doing.

24:53All right, let us review it really

24:56quickly. So, I'm oddly enough I I do not

24:59use the the review uh function. It's

25:03just like not part of my workflow. The

25:05way that I I still use git and like git

25:08diffs to to manage how like how things

25:11have changed. Um so, let us review this

25:15really quickly. So, I'm going to start

25:17with the router. This is the going to be

25:18the most simple. So, okay. So it it

25:22registered the the cats endpoint. So get

25:26cats list. So I'm going to go to the

25:28cats handler and see how it's handling

25:30that. Uh list handles get cats list cats

25:33optional query available for adoption

25:35true

25:38accepts true or accepts one. It goes and

25:42sends that to the list service which

25:47takes in the available for adoption

25:48which then goes and queries from the

25:50repository fetch cats. So cool. So this

25:52this again did what I wanted it to do.

25:55So I'm going to go ahead and uh

25:59uh commit this. You know what I actually

26:01so actually before I do that there's

26:03another general piece of advice that I

26:06have for uh for anytime you're you know

26:09integrating AI into your workflow. Um I

26:12very much think that the the need of

26:15having automated tests inside of your

26:18codebase has increased and it has become

26:21more important um for two reasons. one

26:24because you want to make sure that the

26:26contracts remain stable and you know a

26:28future AI prompt doesn't break anything.

26:31Um, but two, when you actually have

26:34support for uh you you have automated

26:37tests inside of your codebase, the when

26:39you I'm prompting to make a change

26:42inside of my, you know, my cursor uh

26:45window here and I can go and say, hey,

26:46make this change, it can actually go and

26:49run the automated tests uh inside of

26:52this context window and it can go and

26:55verify itself that it did the thing

26:57correctly. Um, so the addition one

27:00additional thing that I'm going to do is

27:01I'm going to have this write some tests.

27:05All right. Uh, so I'm going to say, can

27:07you write some tests? And I already have

27:11a uh I already have a test file here

27:14inside of my

27:16uh service test. So can you write some

27:20tests for the fetch all cats

27:23functionality in the service test file?

27:29>> [snorts]

27:29>> So what I'm understanding from your

27:31process so far just comparing it

27:34>> with um less tightly controlled way of

27:38writing which sometimes I do or maybe it

27:41happens more so in in the front end

27:43you're now writing the back end API

27:44endpoints registering services and

27:47repository I feel like um um you started

27:50small but you're tightly in control of

27:52reviewing every step of the way so you

27:54just don't let it off leash um and Then

27:57you're writing tests so to make sure not

27:59only you're reviewing it, but also tests

28:01are coming in as you're building this

28:03for your back end.

28:04>> Mhm. Correct. Yeah. And so and one of

28:07the really great things about that test

28:09part is like so you can see it right

28:10here. It's popped up. It's like okay,

28:12it's so it just wrote those tests and

28:13now it's like I'm going to I want to run

28:15those tests. So it can run those tests

28:17uh and it run in uh sandbox and it gets

28:20the output of those tests and it goes

28:22and sees the result. Let's see. Did we

28:24run into an issue? uh exceeded. So,

28:26okay. So, I think that it found Yeah.

28:29Okay. So, it found a bug in the test

28:31that it had just wrote because it was

28:33able to run it and actually execute that

28:35code. Um, so it's going to rerun those

28:38tests and make sure that it did what it

28:40wanted to do and it did it. So, the

28:42process of me going and saying now write

28:44tests for the thing that you just did,

28:47it h it gives it a little bit of like

28:50reality like it it it forces it to

28:52actually execute the code. it forces it

28:55to write a test with what its

28:56expectations are for its behavior. Um,

28:59and then and then now I've got a test

29:00there that if I make a change in the

29:02future, it can also modify the test and

29:04verify its contract. Um, so yeah, so

29:06your general summary of of of my process

29:09is kind of correct. It's like so I kind

29:10of start with a a like a shorter leash

29:13and I kind of like give it more and more

29:15and more and more. Um, I would say

29:18professionally I I have reached the

29:20point where I I'm not really doing

29:24There's some scenarios where I go and

29:25say, "Okay, do this very specific,

29:27tightly controlled thing." Um, but I

29:29would say more often than not, I've I've

29:31I've learned I can pretty safely give it

29:34a pretty pretty large leash in the code

29:36bases in which I operate professionally.

29:38Um, so uh I for the most part am I'm

29:44probably about at the level of the pro

29:46the prompting that I just provided where

29:48I'm like build this endpoint. Here's

29:50what you need to do. and I provide it

29:52like the product opinions that I want it

29:54to provide and it can go and infer from

29:57the rest of the codebase of how what I'm

29:59asking for it fits into the codebase.

30:02Um, and yeah, and I always add tests

30:04because it gives it a little bit of um

30:06it gives it a little bit more contact

30:07with reality. Um, okay. So, let's just

30:11let's just go ahead. Let's assume now

30:13that everything is right. So, we're

30:15going to add fully add uh the fetch all

30:18cats uh endpoint. Um and then, you know,

30:23I probably would also test this locally

30:26myself. So, I'm going to just like spin

30:27this up very quickly. I have um I I I've

30:31ran this locally. So, I think I have at

30:35least one cat in my database. So, now I

30:37should be able to just go and say fetch

30:39cats. And cool, it returns all an array

30:43of all of the cats inside of it, of

30:45which I only have one right now. Let me

30:46make another one for my other cat, Ray.

30:49And she is a tor.

30:53Uh, let's create a cat for Ray. Uh,

30:56they're litter mates, so they actually

30:57have the same birthday. So now when I

30:59fetch all the cats, I get both Luna and

31:02Ray.

31:04>> She didn't know you're a you're a cat

31:05person.

31:06>> Oh. Oh, I'm a cat person.

31:10Uh, I can show you photos after. Maybe

31:12you can put some up on the screen. No.

31:14Um, but yeah. No, I'm I'm a cat person

31:16for sure. Um, also too, how's the how's

31:19the size of the screen? It's feeling

31:21like really zoomed in for me, but I want

31:23to make sure it's good for good for

31:24>> I think it it's good. It's good. I know

31:27it makes it hard for you to actually see

31:29what you're doing because it's so big,

31:31but uh on on my end, it's it's very

31:33good.

31:33>> Okay, cool. Yeah, I do I do know it's

31:35Oh, sorry. Go ahead.

31:37Um, no, I just wanted to ask a side

31:39question here. Uh, and that was I

31:41realized um you're still using uh the

31:46agent in the agent mode, not the plan

31:48mode. And also you're not like selecting

31:50any specific model. Is it that like for

31:54the start or when you're tightly

31:56focusing or controlling what it is

31:58producing, you're less worried about

32:00like what model or what mode. Is that

32:02intentional or like what is um how

32:05sensitive you are to those kind of

32:07different models?

32:09>> Yeah. Yeah. So, so I I am sensitive to

32:11these models professionally, right? So,

32:13I I'm what I'm asking of of of uh of

32:17these models right now is actually

32:19really really lightweight. It's this is

32:21a really small code base. The questions

32:23that I'm asking it are really small. Um

32:26so, I just have it on auto mode right

32:27now. So, auto mode is basically it'll

32:29like pick the appropriate model. I don't

32:31know how that actually works underneath

32:33of the hood. Um, but it's generally

32:35good. So, it's that I'm on my personal

32:37license right now, so I have $20 worth

32:39of credits per month. Um, so it's just

32:41like it's it's I would say it's better

32:43in terms of like if you're trying to

32:46optimize a $20 per month cursor license.

32:49Um, I tend to do auto, but but yes, so

32:52professionally, I mean, honestly, I'm

32:54almost always using Opus 4.5. Um,

32:57there's really rare scenarios where I

33:00will actually deviate from that. So,

33:02honestly, like my I just like hardcode

33:04this. It's just always Opus 4.5.

33:06>> I see. Okay.

33:08>> Opus 4.6 did actually come out this

33:10week. Um, I g I gave it a spin for one

33:13day and I've actually switched back to

33:14Opus 4.5.

33:16>> Um,

33:17>> yeah, I have found there was I it I

33:21encountered this. So it was having a

33:24hallucination and I could not convince

33:26it out of this hallucination. It was

33:28basically assuming that there was this

33:30bug with my this underlying library that

33:32I was using and it did not exist. I was

33:34like no the bug doesn't exist like write

33:35a test proving that this bug exists and

33:37it couldn't write a test and it's like

33:38yeah but the bug still exists and I was

33:41in Opus 4.5 and I'm like okay I'm kind

33:43of losing my mind right here. Um so I

33:45switched back to and this was only I

33:47used it for just one day. Um and then I

33:48switched back to Opus 4.5 and it

33:50immediately was like oh yeah you're

33:51right. Yep. I can see from the test that

33:54the the thing that I was previously

33:56saying uh was not true. Um so that was

34:00my experience with it. It's probably uh

34:03just like a one-off thing, but uh but

34:05yeah, I'm still on Opus 4.5. I'll

34:07probably give it another spin this week

34:10and see how it is. Um but that was my

34:12initial experience with it. Yeah, there

34:14seems to be like a um

34:18um kind of have like two different sides

34:21of how people are finding 4.6. Some are

34:25just like what you said um staying with

34:28the 4.5. I'm not really surprised by,

34:32you know, 4.6 or like the people who are

34:34I I see it on Twitter. They're like,

34:36"Okay, what have you done with the 4.6

34:37that's just like blew your socks off or

34:40how is it different?" So it seems like

34:41the 4.5 is still um kind of the safe

34:44zone and trusted model for

34:46>> Yeah. Yeah. I mean definitely like so

34:49the jump from preopus 4.5 to Opus 4.5

34:53was like holy Um the jump from

34:56Opus 4.5 to Opus 4.6 so far has been

35:00like I beyond that one negative

35:02experience that I had with it. It hasn't

35:04felt like profoundly different. Um, you

35:07know, again, maybe it is contextual to

35:10the code bases in which I'm operating.

35:12Um, and maybe there's other scenarios

35:14which just like much better,

35:17but yeah, that was that was basically my

35:19my experience with it. It wasn't

35:21significantly improved. So, I'm still on

35:23Opus 4.5.

35:24>> Okay.

35:25>> Um,

35:25>> and if when you're using these models,

35:27are you um like professionally using it

35:29within cursor or do you use like cloud

35:32code to just look at terminal based?

35:34>> It depend. It depends on the task. Um,

35:37so I do I I still do like cursor. I like

35:39I like cursor because it it still gives

35:41me that like interf I've been coding

35:45well I used to write code like a you

35:48know we used to and like by hand. Um I

35:51like cursor because it gives me the it

35:54gives me this a similar experience that

35:56I'm used to which is probably this is

35:58like a very much like a legacy you know

36:01legacy developer opinion. Um, so I can

36:04still go and I could jump into files. I

36:06can say really specifically highlight

36:08code say, "Oh, this is I want you to do

36:10this very specific thing. My my prompts

36:13are really really specific often because

36:16I have opinions and I want those

36:18opinions to be reflected in the co in

36:20the output, right? I like to have the

36:21code that AI writes virtually be

36:24identical to the code that I would

36:25write. I like to look at it and say,

36:27"Yep, that's how I would have written

36:28it." And I like that to kind of be the

36:31standard of the quality of the code that

36:33I that is output with when I'm using AI.

36:36Um,

36:37>> I think that that is a not necess it's

36:40not necessary in the same way in and

36:43it's still a thing that I'm getting used

36:45to, but it's like

36:48AI there's in in the old world you used

36:52to encounter things where you're like,

36:53"Oh, there's this issue. It's not really

36:56great. Let's like put it on the

36:57backlog." And you like never would get

36:58to it. But in the new world, you're

37:00like, "Just fire off an agent, man.

37:02It'll fix it. it'll improve the thing

37:03that you wanted to do. Um, and so for

37:06those scenarios, it's like that that's

37:07when I'm I'm using more like clawed is

37:09is when I want to have like an agent. I

37:11use I'm a I used to fully operate inside

37:14of my terminal. So I'm a big Vim user.

37:16Um, uh, I so I use like T-m so I just

37:19like spin up a new T-Max pane and and

37:22fire off an agent and and you know let

37:24it go to town. Um, and it's generally

37:26quite good at that. So I would say it

37:30professionally I would say it's it's

37:32more varied than just cursor. Um you

37:35know there's times though I but I would

37:38say probably cursor is my daily driver

37:39right because it gives me that

37:41experience that I'm used to. Um maybe

37:43I'll shake that when I'm like you know

37:45I'm not really doing any you know

37:48jumping jumping in and editing things

37:50myself and I'll just fully switch over

37:52to cloud code. Um but uh but yeah,

37:55that's generally my experience. And then

37:56so to to go back on a question that you

37:58had asked earlier, um oh man, I'm going

38:01to give you such a fun time in the

38:02editing room, by the way. Like

38:04[laughter]

38:05it's you're going to have to cut you're

38:06going to have to cut this down in a in a

38:08fun way, I presume, unless you want this

38:10to be a full hour. Um um sorry. Uh so to

38:14to go back to uh you you asked a

38:16question about plan mode. Um so I do use

38:18plan mode. Um I haven't I haven't

38:21reached that yet inside of my um uh so

38:24basically so I've reached plan mode when

38:26I am not confident that it's going to be

38:29able to come up with the right plan in

38:33the first shot. Right? So the thing that

38:34I just asked it of I know it's going to

38:36be able to do it because I you you kind

38:37of you gain an intuition of what AI can

38:40do the the more that you use it. Right?

38:42And so my intuition of this codebase is

38:45I I I probably don't need to use plan.

38:48Um maybe we maybe we will we'll we'll

38:50we'll do a plan we'll do a plan thing at

38:52the end. We'll give it a really big I'm

38:53basically going to tell it build the

38:55rest of this API right at the at the

38:56very end because we're not going to be

38:57able to step through every single

38:59endpoint. We'll go through a plan mode

39:01and we'll make sure that hey is this is

39:02this going about it the the right way.

39:04Um so we will use that and yeah I use

39:06plan mode when I don't trust that it's

39:08going to come up with a good plan and I

39:10want to see what it plan is before it

39:11goes and executes. But for a lot of

39:14tasks, it's like, I know it's going to

39:15do this right because I've given it a

39:17reasonable enough scope and it's going

39:18to be able to execute based off my

39:20prompt alone. And worst case scenario,

39:22it does it wrong. I click reject and

39:24then I go into plan mode. Um, so yeah,

39:26plan mode is a really valuable tool. Um,

39:29another workflow that that one of my

39:31colleagues uh uh just just introduced me

39:33to is rather than actually using plan

39:37mode, create a file for that for a plan

39:41and actually use uh just use the the uh

39:45uh agent mode [snorts] and have it write

39:48the plan to the file. And that makes it

39:50easier where you can jump in and you can

39:52actually go and edit that file. Plan

39:53mode is like to to edit a plan I've

39:56found it's a little bit more clunky. I

39:57have to prompt it in particular way to

39:59go and edit that plan. Um, but having an

40:01actual file is really nice because you

40:03can actually if you decide you want to,

40:05you can keep that as an artifact and

40:07actually commit that plan so people can

40:09understand the context and the reasoning

40:11behind your decisions in the future. Um,

40:13I think this is going to be a more of a

40:16I think that committing the context and

40:19the reasoning and kind of some artifact

40:22of the prompts I think is going to be a

40:24thing that is going to occur in the

40:27future of like you know AI native

40:29companies like I think it's going to be

40:31like yeah like write this write this PRD

40:34for what the feature that you want to

40:36build out um and you know work with AI

40:40to build out that PRD and and point AI

40:43to that BRD and say, "Okay, now go and

40:45build it." And then also to committing

40:47that PRD because it's it's this really

40:49useful artifact of like, "Oh yeah,

40:51here's the rationale behind all of my

40:52decisions." Cuz sometimes the rationale

40:55behind your decisions, it's lost in this

40:57in this context thread and it's maybe

40:59not reflected inside of your codebase.

41:02>> Yeah, I think that you nailed it. I I my

41:04takeaway from what you said is two

41:06things. one once you start building

41:09different features inside of a codebase

41:10you develop this intuition on what the

41:13agent can do what what they're good at

41:15what they're not so you know where you

41:17can trust it and what not and the second

41:19thing is what you mentioned about like

41:21the this PRD document that you create

41:23this instead of like having using to

41:25plan mode maybe use the agent build that

41:27PRD go in and just make it as tight as

41:30you want then point DLM at it and see

41:32like this is this is a context go build

41:34this now I think I've tried it a a

41:36couple times and I think that's that's a

41:38better better way of using a like um

41:42kind of established plan rather than

41:44just using the plan mode. Yeah. Yeah. So

41:47so yeah. So anyways, it's it's it's also

41:49too like so I think that like we I think

41:51we still there's it's it's this video is

41:55going to be stale in two months because

41:56it's like the the way in which we are

41:58using AI has is changing so drastically.

42:00like I add some new tool to my to my

42:03workflow or add a new way that I'm using

42:06uh AI every single week it feels like.

42:10Um, so it's like there's it's there's

42:12it's it's it's it's

42:15in every regard it's software

42:16development accelerated, right? The the

42:19history of writing code has always been

42:21I've got to learn a new framework. I've

42:23got to learn a new language. I've got to

42:24learn a new thing. You've you always had

42:26to learn, but it was like the cycle was

42:28like a year. And the the world that

42:30we're in now, the cycle is like it's

42:32like every single week it's like, oh my

42:34god, what like what the what is what is

42:36what is Clawbot like? Oh, it's not

42:37clawbot any like it's like anyways it's

42:39there's there's this really really fast

42:41iteration that's happening right now

42:43which is like I think from a

42:45>> I think if you have the right approach

42:47and philosophy it's like you you very

42:49much if you go into this has always been

42:51true if you go into and you you're a

42:53software engineer you want you want to

42:54go into software and you want to write

42:55code you have to have a learning mindset

42:58the entire time throughout your career

43:00there has not been a month in my career

43:02where I have not encountered a problem

43:04where I had to learn a drastically new

43:07thing to be able to solve whatever

43:09problem I'm solving or you know the the

43:12world has changed and now we're using

43:14this new framework. Um you always

43:17constantly have to learn. I think the

43:19that has accelerated more. So you really

43:21have to very much accept like you know I

43:24I don't know what I don't know and I'm I

43:27you need to have to kind of really

43:28really be willing to constantly iterate

43:30and constantly like learn new things.

43:32Um, and yeah, and it's just it's just

43:34more accelerated. And that's true for

43:36everything about AI. It's just more

43:37accelerated. It's more of the same, just

43:39accelerated.

43:40>> Yeah. Yeah. That's so true. It may be

43:42harder to keep up at the speed at at the

43:46speed at which it's changing, but it's

43:48at the very core you're learning, um,

43:50you know, problem solving with new new

43:53tools now.

43:54>> Yeah. Yeah. Exactly. Um, cool. Okay. I

43:56want to do a plan I want to do a plan

43:58mode now just so we can kind of like

43:59walk through that workflow. Um, and then

44:02we can talk. I don't think I'm I don't

44:04think I'm going to I mean I could I

44:05could build out a UI so we can look at

44:07the UI and just give it give a once

44:10maybe when we're talking I'll give it a

44:11prompt. Okay, now build a UI that

44:13interacts with this API and we'll see

44:14how that works. Um, all right, cool.

44:16Okay, so I'm going to yeah, I'm going to

44:18give it I'm going to give it a plan

44:19mode. Um, and I'm going to say um let me

44:23close everything here to make it nice

44:26and clean.

44:27So, oh, let me just add this check mark

44:30here just so it knows. Um, okay. So, I'm

44:35going to I'm going to use this same

44:37thread because it's probably got some

44:38good context inside of it. So, okay. So,

44:42I now want you to build

44:46all outstanding endpoints

44:50as described

44:53in the readme file.

44:56are not yet implemented.

45:00Actually, that's what I'm just going to

45:02do that. I'm not going to give it more.

45:04I have a tendency to really really write

45:06a lot in my prompts, but it you often

45:08would be surprised. And you know what?

45:10Hold on. I'm going to stop this really

45:11quickly. We are going to

45:13>> We're going to use Opus 45. Uh

45:18uh add models. Let me just add Opus 45

45:21really quickly. We could pretend that we

45:23recorded this video before Opus 4. No, I

45:25was talking about Opus 46. All right,

45:27Opus 45 brain mode. So, we're pulling

45:31out the big the big guns here, right?

45:33We've uh we're get we're we're asking it

45:36a really big task. So, we want to use a

45:39really smart model and we did the

45:41>> with the brain icon which is

45:44>> uh thinking which I think that's

45:45actually I think that's the the the way

45:47in which uh cursor actually provides the

45:50context and it operates. I think that's

45:52a harness function and not like a

45:54different model. Um so okay so cool. All

45:57right so open 45 go ahead and build it.

46:00Okay cool. While this is working we can

46:01we could probably we could probably talk

46:04um about other things. So yeah

46:06>> so one thing I wanted to ask is like now

46:09we are doing that same kind of idea

46:11where that your readme file is like a

46:13PRD document. It's just like what I want

46:14to build it and it has enough context in

46:17what has already been done. So you can

46:18go and check those references and then

46:20kind of build the rest. I and we're also

46:23using the plan mode. So it's just like

46:25um doubled up the layer of planning with

46:28a PRD and also plan mode in cursor. Here

46:31I also see like there is an

46:32architecture.md file in in your uh

46:36source code. Is there any other context

46:38you're providing to the LLM?

46:41>> Uh so in this codebase only the

46:44architecture.md file. So I think I I

46:47think it uh I don't even know if it is

46:49reading from this file in this thread.

46:51Um I I bet it I bet it encountered this.

46:54Um but basically so yeah so basically

46:56what this architecture file is is is

46:58what I was kind of describing earlier.

46:59It's like I'm providing like the

47:01highlevel architecture of of of how I

47:05want this uh API to be structured. And

47:07so specifically this API it's like

47:09layered architecture. It's like you know

47:11I've got my at my lowest level I've got

47:14my database that stores data. Then I've

47:15got the repository layer which is the

47:17abstraction that defines how you

47:19interface with it. Then I have my

47:20service layer which can provide

47:23additional functionality on top of that

47:24and then the HTTP layer which is

47:27ultimately what the end consumer of this

47:29API is going to be interfacing with. So

47:32I have some some highlevel architectural

47:35here are the layers and also too for

47:36what it's worth this is this file I I I

47:40gave it the highle context written in

47:42really sloppy English and I had AI write

47:44this right and I verified it I made a

47:47few edits made sure that it was correct

47:49um but it was able to kind of actually

47:51go and you know write this itself so you

47:54don't have to put a lot of effort into

47:55these files but they're actually really

47:57helpful because what you can do is again

48:00like mentioning kind of like how you you

48:02you knew somebody who has like a million

48:04line codebase or whatever. It's like you

48:06don't want the agent to have to do that

48:08every single time. Go and read the

48:10entire codebase to figure out okay

48:11what's the architecture of this thing.

48:13You kind of can provide these like

48:15summarized bits of context. You can do

48:17the work up front provide that summary

48:19and then drop that summary somewhere and

48:21then it's you know a useful bit of

48:22context that can be provided to use in

48:24like other future prompts. Um, you could

48:26use that in your agents.mmd file, which

48:28I believe is automatically pulled into

48:29basically every prompt context inside of

48:32um, cursor. Um, or you can have kind of

48:35more oneoff specific things where you're

48:37like, ah, there's a specific I want to

48:39pull this in because it's relevant and

48:41you can kind of have different markdown

48:42files throughout your codebase.

48:44>> Yeah, I really like this idea of because

48:47as you mentioned, agents MD is just

48:49automatically pulled in. So if you just

48:51bloat that context is not necessarily

48:53always good. But if you're having

48:55multiple different files and put a

48:57reference inside of your agents.mmd to

49:00say hey if you want to learn about

49:01architecture here's here's where it

49:03lives. So it doesn't

49:04>> it pulls it as it needs. So this is how

49:07I do typescript or these are my types

49:09these are my I don't know different

49:11layers and you have like these different

49:13multiple files but which all of them are

49:16referenced in your agents.mmd. So it's

49:18just best of both worlds. So you're not

49:20bloating your context, but you're also

49:22giving it access to specific M markdown

49:25files when it needs to pull pull from

49:27them.

49:28>> Cool. Yeah, that that's a smart idea. Um

49:30I think there's there's repos at work

49:31where we have setups like this. Um I

49:34haven't set that up in my personal

49:35projects. I totally should because it's

49:37like Yeah, cuz I I I I think it's it's

49:39probably a reflection of how I actually

49:41do a lot of prompting. I'm I'm really I

49:43like to participate. You don't see me

49:45build do a lot of prompting like this

49:47where I'm like just do the thing. I'm

49:48like, do the thing in this particular

49:50way. As such, I often don't find myself

49:53needing to have context be automatically

49:55filled because I am often the one that's

49:58providing the context. I think this is a

50:00this is totally like an old way of

50:02operating and it's like I' I'm

50:04definitely shaking it off and I'm kind

50:06of I'm switching into trusting the the

50:09AI agents more and more to go and say,

50:11you know, things like, you know, kicking

50:12off background agents and stuff like

50:14that. That feels like a It feels like a

50:16Hail Mary sometimes when you're like,

50:17can you try to do this thing? I'm not

50:19going to even be participated

50:20participate. Just like submit a PR once

50:22you're done. Um that's definitely uh

50:25kind of like the next layer for sure. Um

50:27okay, but it's it's got a plan. So I

50:29think let's let's review this plan

50:30really quickly before we um before we

50:32continue and then we can press build and

50:34then we can we can go back to chatting.

50:36Um okay, so implement all the

50:38outstanding endpoints. Okay, so there's

50:3914 outstanding endpoints across four

50:41resource types. Here's the architecture

50:43pattern.

50:45Great.

50:48Yep. List all customers. Delete

50:50customer. Looks good.

50:53Delete.

50:56Okay. Update cat. Delete a cat. Yeah.

50:59Cool. Okay. It it appears this isn't a

51:01super complicated codebase. So it I

51:03think it came to a pretty good

51:04conclusion on how it should go and build

51:06these things. Um a lot of my So this is

51:09okay. This is another very small thing.

51:11But so one thing that I did if you're

51:13building a backend API um a lot of your

51:17product decisions come out when you're

51:19building out the database schema and

51:21then everything else on top of that is

51:23just you could largely describe it as

51:25boilerplate. Um so I I build my schemas

51:30like I'm writing my database migrations

51:33myself still like I sometimes I you know

51:36write a prompt to make it like easier

51:38here and there. Um, but there's a lot of

51:40product decisions that come, right? So,

51:42for example, here's I've got

51:44reservations. I have reservations where

51:46it's like every cat can only have one

51:49reservation per hour on and every

51:51reservation is on the hour, right? These

51:53are product decisions that AI can

51:56actually go and look at my database

51:57schema and go and say, "Oh, okay. I

51:59understand this product decision and

52:01actually design the correct API to to

52:04satisfy your requirements." So I build a

52:07lot of I still build my migrations and I

52:09still define the underlying database

52:11models um because when you're building

52:12an API you very often uh are able to uh

52:17it's able to infer these things and

52:19build and build it correctly um from

52:21that um plugin.

52:22>> True.

52:23>> Um okay cool. All right. So now we're

52:26building an entire API

52:29and then so what and then so after after

52:32this two I can send off another prompt

52:34and say just build a UI that goes and

52:36queries from this and we'll we'll see

52:38how it does. Um cool okay while this is

52:41going we'll we'll we'll let this just go

52:43and uh chat more. I think this idea that

52:47you mentioned uh starting from your you

52:50know database schema as the font

52:53foundational level for what everything

52:56else is just kind of build up build on

52:58top of um it makes um a lot of sense.

53:02you can maybe collaborate with uh AI to

53:04build that really go back and forth, but

53:07that that's where you need to probably

53:09spend the most amount of time, you know,

53:11thinking about the use case or the the

53:13business or the service layer or

53:15whatever it is that you're building and

53:17then going back into your data schemas.

53:19I think everything else just really sits

53:22on that foundation.

53:24>> Yeah. Yeah. Exactly. Um this this is

53:27this is how it's always been, right?

53:28It's every everything I I very much like

53:31I have the philosophy your database is

53:33your application um because it's the

53:35thing that breathes life into your

53:36application. It's the thing that you

53:38know persists from from time to time. So

53:40it's like yeah you it's it's a really

53:42consequential thing and and and a lot of

53:44the product decisions that you make come

53:46from this. So it's it's it's definitely

53:48a great place to like if you're going to

53:50be involved and you're building a

53:52backend be involved when you're

53:54designing out your schemas. um the APIs.

53:57Okay, you can go and do this. You could

54:00just have cursor, you know, send a

54:03prompt, build all these endpoints. It'll

54:05probably do a reasonable job. Um so,

54:08okay. So, it's still doing this, but

54:10yeah, so this is definitely a great this

54:11is definitely a great place to if you're

54:13going to participate,

54:15make sure you don't skimp out on on

54:17this. Make sure that this is in the in

54:19the right way that you want it to be.

54:25Now I know um at some point we're going

54:27to also use coding review agents and

54:32services to also look at what we have

54:34produced

54:36>> um from the back end to the front end

54:38>> uh as well. Maybe while this is building

54:40we can also talk about that part as

54:42well.

54:43>> Yeah totally. So this is definitely one

54:45part of the stack that is the AI uh

54:48stack that is very uh I think very

54:52important. Um, so one of the downsides

54:54about uh you know uh like uh AI written

54:59code is it can sometimes have bugs like

55:01this is just kind of like the reality.

55:03Um it can introduce bugs into your

55:05codebase. Uh and that is you know if

55:09you're running like a production grade

55:10application that can be a big problem.

55:12Um so we so I I use uh at work and and

55:16also to personally uh I use a tool

55:18called code rabbit which is basically a

55:22uh uh a tool that will automatically

55:25review your code um when you submit a PR

55:27to GitHub. Um this is actually this will

55:30be a great uh this will we'll make a

55:32branch and we'll submit a PR and we'll

55:33we'll let code rabbit run because this

55:35is we just gave AI like a lot of

55:36responsibility in this PR like build

55:38this whole thing. Um, so we'll see maybe

55:40if Code Rabbit can catch anything that

55:42could be improved. Um, but yeah, it's

55:44it's a really really useful part like

55:47especially so if you're at a company you

55:48haven't adopted AI yet, this is a great

55:51place to adopt it because it's not going

55:53to introduce new bugs into your

55:55codebase. is going to prevent you

55:56prevent bugs because I think that a lot

55:58of the reservation trying behind where

56:01some companies haven't quite introduced

56:03AI into their workflow is because of

56:04this resistance of some nervousness of

56:06like well if I give up you know this

56:09responsibility to the robots like what

56:11if they make a mistake like how will

56:13they you know who's responsible for that

56:16etc etc so I think that concern is it's

56:20it's actually the opposite with like an

56:22AI code review tool it's like no it's

56:24it's reviewing ing your code and it's

56:26it's it's actually it's it's

56:27surprisingly good actually. It's way

56:30better at reviewing code than I am

56:31personally. Like I can review code from

56:33a from a [snorts] design how you've

56:35designed your made your highlevel

56:38decisions. Um but yeah, like line by

56:40line code reviewing like catching like

56:42oh you've got a one-off error here or

56:44like oh you're using this this uh you

56:46know API incorrectly or something. Um

56:49I'm quite bad at you know because it it

56:52just requires like such a lot of effort.

56:54Um, okay. It's asking to uh run some

56:57tests here. So, I'm going to let it run

56:58some tests. Um, but yeah. So, so yeah,

57:01so we will we will we will submit this

57:04uh through uh a PR and I have I have on

57:06this repo I have Code Rabbit wired up.

57:08Um, so we'll see what it can do.

57:11>> Beautiful.

57:11>> Um, okay, cool. It says build it. Um,

57:15we're not going to review it. We're

57:16we're in, you know, ourselves. We're not

57:19going to review it. Okay, keep all Yeah,

57:21it got it got So, see, look at that. key

57:23business rules implemented. So, as I had

57:25mentioned, it's important to write

57:27inside of your database schema. A lot of

57:28your like, you know, you make a lot of

57:30the product decisions in the business

57:31rules. This was a this was like a

57:33business rule decision that I had

57:35written into my schema and it was able

57:37to infer this and and and apply that.

57:39>> Um, so, okay. So, cool. So, I'm going to

57:42we're going to add this. Oh, we're going

57:43to go to a branch. Uh, we're just going

57:45to call this full API.

57:51All right. Let's just add Okay. Uh

57:55uh adds the full API now. Adds

58:06uh full API.

58:09Okay. So, cool. So, let's go over to

58:14let's go over to the repo here. Um, let

58:17me let me move this over and pull this

58:21over here by itself. And let me close

58:24this down.

58:26[sighs]

58:27Okay, cool. So, let's go over to the

58:29repo here and let us uh let's submit a

58:32pull request here. Um, this is exactly

58:34the same workflow that you would

58:36normally use on a team. Um, so I'm going

58:39to submit this pull request here. Uh and

58:42look, yeah, this is uh this is a lot of

58:44code that if I submitted a PR like this

58:46to uh to one of my co-workers, they

58:49would be like, can you please break this

58:51down into multiple uh uh you know, PRs

58:54because it's so big. But now, because we

58:56have we have AI code review, the risk of

59:00a PR this large has gone down quite a

59:02bit because it's like all those like

59:03actual like bugs that like might pop up

59:05and might manifest are uh are reduced

59:09because it's going to go and check out

59:10the PR. So, look. So, okay. So, this

59:12cute little rabbit here, uh,

59:16goes and says, uh, okay, I'm going to

59:17review your PR. Uh, it might take a few

59:19minutes. It usually doesn't take a few

59:20minutes. It's usually actually very

59:22quick. Um, and, uh, if it catches any

59:25issues, it's going to add a comment

59:27inside of this PR um, and say, "Hey,

59:30have you considered this? You know,

59:32maybe, oh, this is looks like a bug."

59:34Hopefully, there's actually a bug. If

59:35there's not a bug, we can force a bug. I

59:37can introduce like a one-off error or

59:39something like that.

59:40Um

59:42and uh we'll see if we'll see if this uh

59:44catches it.

59:45>> Catches it. Yep.

59:46>> Yeah. Yeah. But but yeah, no, Code

59:49Rabbit rabbit code rabbit is really

59:50great. I um I'm uh definitely pretty

59:52friendly with the team there. Uh it's a

59:54super super talented team that's

59:55building out this product. Um the thing

59:57that's really great about Code Rabbit is

59:59they are building out a it's like a very

1:00:03focused tool, right? So you have other

1:00:05tools like like uh you know cursor has

1:00:07bugbot and you know GitHub has their

1:00:09their you know one tool uh with uh uh

1:00:13what is it called whatever their code

1:00:15review tool is called. Um anyways

1:00:17similar functionality. Yeah.

1:00:18>> Yeah. Yeah. Similar functionality. I

1:00:20really like how code rappers are really

1:00:21really focused on this one particular

1:00:24business concern. Um so it actually is

1:00:26like a really really really solid

1:00:28product. Um, so we'll see if it if it

1:00:31can uh if it can prove me right though

1:00:33and catch catch a bug. Hopefully AI

1:00:35introduced a bug and Code Rabbit can can

1:00:37do something. But it's this is such a

1:00:40simple thing I'm asking AI to do like

1:00:41these endpoints that I'm asking. It's a

1:00:43it's a green it's a green field

1:00:45codebase. It's not really simple

1:00:48complicated business concern like I'm

1:00:50willing to bet there might be no bugs

1:00:51inside of this PR. Um, so we might have

1:00:54to force one but we'll see. We'll see.

1:00:56We'll see how it goes. But I like the

1:00:58idea that you mentioned it's a very good

1:01:01entry point for businesses that are

1:01:03still skeptical um to use AI in that it

1:01:07is just it can also review codes that

1:01:11regular engineers are writing if anybody

1:01:14is writing code by hand anymore. But um

1:01:17it can sit on top of what you're doing

1:01:19at the end of the stack to just review

1:01:21the PR codes and especially when there's

1:01:23like big enough changes. Um this goes

1:01:27line by line and the level of detail

1:01:30that it can have the review on your

1:01:32codebase. Usually no

1:01:35uh human being does that level of code

1:01:37reviews. It's just more so high level or

1:01:40the PR needs to be so small that can be

1:01:42manageable to review it or it's just

1:01:44like very high level. But this thing can

1:01:46just go inside.

1:01:48>> How long to

1:01:50>> Oh. Oh, sorry. Go ahead.

1:01:52>> Um, no. I I was just gonna ask how long

1:01:54would it usually take to

1:01:56>> to kind of review a Yeah. So, this so

1:01:58this is a bigger PR. So, this might take

1:02:015 minutes. Um, but so it's certainly

1:02:03going to be way faster than a human can

1:02:05get to actually go and review your PR.

1:02:07Um, so but so one additional thing that

1:02:09that I'll like to say about this is kind

1:02:11of to add on to what you've said. This

1:02:13is an additive thing to a human a human

1:02:16review, right? If you're working on a

1:02:18team, um you you still probably do want

1:02:20a human review to approve before you

1:02:23know you you you you go and say, "Okay,

1:02:25let's merge this in." Um that's

1:02:27certainly what we what we do at Clerk.

1:02:29We don't go and say, "Oh, okay, let's go

1:02:30and merge this. Let's go and merge this

1:02:33in just because code rabbit thought it

1:02:35was good." Um it's a fully additive

1:02:37thing that can catch things that the

1:02:39human might not catch. Um, and kind of

1:02:42as you mentioned, yeah, it's it does

1:02:43that line by line. It's very distrustful

1:02:47of your code. It doesn't like when when

1:02:48a human's reviewing your code, I would

1:02:50say you often kind of take like I trust

1:02:52that this person is using these

1:02:54interfaces correctly. Like maybe that's

1:02:56like maybe the wrong posture to take,

1:02:58but it's like you know if you get a if

1:03:00you get a you know uh a 1900 line note

1:03:05like it's basically a 2000 2,000 line

1:03:07decode PR you know you're going to have

1:03:10to you're going to have to skim to some

1:03:13extent on some things like you're not

1:03:14going to be able to review every single

1:03:15line be like oh yeah how does this how

1:03:17does this specific function work and you

1:03:19have to jump say oh how does this

1:03:20function work like but code rabbit goes

1:03:22and does that it will actually trace

1:03:24basically every function function call

1:03:26that you have and it will be like oh

1:03:27yeah you're using this API correctly or

1:03:29oh no you're using this API incorrectly

1:03:31um so uh yeah it's it's definitely like

1:03:35an additive thing right don't treat it

1:03:37as like it's a replacement for human

1:03:38reviews like I think that makes I think

1:03:40that makes people especially teams who

1:03:42are nervous about AI in general um

1:03:45that's not the go that's not the goal of

1:03:46a tool like this tool like this is like

1:03:48to help you catch additional things and

1:03:50this and this this tool does help you

1:03:53catch bugs from production like we There

1:03:55are there are actual bugs that like I've

1:03:58gotten human approvals and it goes and

1:04:00says, "Oh, there's an issue." And it

1:04:02catches a and it catches a thing. Um

1:04:04like this happens quite often. Um it's a

1:04:07really really invaluable part of our

1:04:08process. Like we I'm I'm a huge fan of

1:04:11Code Rabbit.

1:04:12>> Beautiful. Yeah, they're um they're

1:04:14kindly respons uh sponsoring this video

1:04:16as well. So there's a link down in the

1:04:18description if anyone wants to check

1:04:19them out. Uh, we're going to have

1:04:22hopefully more content about how to

1:04:24integrate this into your workflow and

1:04:26use it um maybe with Brandon in future

1:04:30again, but definitely check them out.

1:04:33Awesome. Um, yeah, cool. Okay, so we

1:04:35just got we just got the response here

1:04:36from from the code rabbit review. Um,

1:04:39and okay, it's it found a few things.

1:04:40So, let's see what it found. Um, so

1:04:43okay, so a non-atomic multi-write. Um,

1:04:46so I think that uh basically we're not

1:04:48doing something in inside of a

1:04:50transaction here. So if for whatever

1:04:53reason your uh I think it's in the it

1:04:56appears to be in the adoption service.

1:04:58Maybe this is like when it's maybe it's

1:05:00like when the adoption thing is

1:05:01approved. Um and you're and at the same

1:05:03time presumably we're supposed to

1:05:05transfer the owner of the cat to the

1:05:06person. Um this can fail leaving a cat

1:05:09in limbo which would be so say it they

1:05:11would be marked as adopted but not have

1:05:13an owner. Um, so this is a major issue.

1:05:15Um, so, you know, I would actually go

1:05:17and say, "Oh, okay. Let's go let's go

1:05:20and fix this." Um, let's see here. So,

1:05:22birthday clearing uh is silently

1:05:25ignored. Um, so yeah, here's probably a

1:05:28pretty legitimate issue. Um, and then uh

1:05:32here's uh another one. Let's see what

1:05:34this is. Struck comment misleads about

1:05:35null. Um, yeah. Okay. So, like it needs

1:05:38to be it needs a little bit more like

1:05:41clarifying comments. So here's here's

1:05:43one last thing that I'll show. I'm not

1:05:45going to fix all of these because you

1:05:46know we're not going to this is not a

1:05:48real real codebase. This is just a fun

1:05:50codebase. But like these are actual real

1:05:52things that you would want to fix if you

1:05:53were in a real codebase because even

1:05:56even though the the the issues that you

1:05:58think well ah that will that case will

1:06:00never happen like you know a transaction

1:06:02ah do we actually really need a

1:06:03transaction? What are the chances that

1:06:04the database fails on the second thing?

1:06:07um when you're operating at scale in a

1:06:08real codebase and you have you know uh

1:06:10you know 10,000 requests per second to

1:06:12your API um edge cases happen all the

1:06:15time so you would actually want to fix

1:06:17this. So, here's one of the great things

1:06:19about Code Rabbit. Um, so I'm using

1:06:21cursor. Um, I can code rabbit provides a

1:06:24prompt that I can go over to cursor and

1:06:28let's see. Oh, here we are. Um, I can go

1:06:30go over to cursor. So, I can just go and

1:06:31say copy this prompt right here. Uh, and

1:06:34I can go and just drop this in here. I'm

1:06:37not even going to review it and let's

1:06:39see what happens. Um,

1:06:42I really like that feature. So it

1:06:44plugged out the issue, gave you a

1:06:47prompt, copy pasteed in your agent, it

1:06:50just pointed at the place goes and fixes

1:06:53that with enough context.

1:06:55>> Yeah. Yeah. And then so one so we'll

1:06:57we'll see. It's pro I think it's

1:06:59probably going to be able to fix it. I

1:07:00have a way to create a transaction

1:07:01inside of this codebase. Um,

1:07:05>> one thing I want to ask people while

1:07:06this is doing like how um, sensitive are

1:07:11you in checking your context current

1:07:13context window? I know cursor has this

1:07:16thing that you could turn on that it

1:07:17shows how much of the context it has

1:07:18already used like there's like a little

1:07:21radar there. Do you check that? Are you

1:07:24worried about like if you're blowing out

1:07:26the context or cuz you just keep pasting

1:07:28it in the same kind of dialogue that you

1:07:30have because it has relevant information

1:07:32in in what it has already done. Uh do

1:07:35you ever start new chats or if yes at

1:07:39what point?

1:07:40>> Yeah. So I I usually so I usually create

1:07:43a new chat every time I move into a new

1:07:45problem. Right. So it's like it's almost

1:07:47like one to one. We use a we use linear

1:07:48for our issue tracking system. um it's

1:07:51almost one one to one when I'm working

1:07:53on you know build this particular

1:07:55feature and we have a linear ticket for

1:07:57it I try to keep the same chat window

1:07:59because there's usually relevant context

1:08:02in either my prompts or what it has

1:08:04fetched before um so I'm usually not too

1:08:07terribly concerned about it the the

1:08:10um like I I do keep an eye on like what

1:08:13my costs are inside of C like I I can

1:08:15access my dashboard um the ROI on any

1:08:18given prompt for my business is it's

1:08:22it's it's not I I never encounter a

1:08:24scenario where it's like, "Oh my god,

1:08:25I'm like spending too much on tokens."

1:08:27Um uh so it's not a thing that I super

1:08:30concern myself with, but I do sometimes

1:08:33want to clear the context because I've

1:08:35moved on to a new problem and I don't

1:08:36want it to get confused that it's still

1:08:38working on the same problem. Sometimes I

1:08:39have had that happened where I moved on

1:08:41to a new problem and I haven't changed

1:08:43the window because I forgot to change

1:08:44the window. Um, specifically because we

1:08:47are just like working on like we're like

1:08:48building the API is like I would say

1:08:50what the task is like the initial build

1:08:52out of the API. I haven't changed the

1:08:53window because I think there's probably

1:08:55valuable stuff up there.

1:08:56>> Yes, makes sense.

1:08:58>> Yeah. Um, cool. Okay. So, let's see what

1:09:00it went and did. Um, so, okay. So, I

1:09:02think now it's making a uh it's

1:09:06beginning a transaction

1:09:11and it's committing the transaction at

1:09:12the end. Cool. So, let's go and add this

1:09:15um uh fix uh atomic

1:09:20uh transaction issue. This I'm just

1:09:23going to say fix transaction issue. Um

1:09:26look at me worrying about my commit

1:09:27messages on a project that does not

1:09:28matter. Um I cannot I cannot train it

1:09:32out of myself. [laughter]

1:09:34Uh okay.

1:09:37Uh okay. So, one of the great things

1:09:39about code rabbit as well is it's like

1:09:41okay cool. you have this issue,

1:09:44Code Rabbit's going to see that new

1:09:45commit come through. It's going to say,

1:09:46"Oh, you fixed that issue and it will

1:09:49resolve this conversation." Just like if

1:09:50you had a human reviewer submit a PR and

1:09:52they say, "I want you to fix XX and X."

1:09:54You submit some commits. You reerequest

1:09:56a review. The usually the person who

1:09:59submitted the concerns go and say, "Yep,

1:10:00you resolved you addressed my issue. I'm

1:10:02going to resolve this comment." Um, so

1:10:04it's exactly the same flow in in code

1:10:06rabbit as well. Um, so the great thing

1:10:09is that you do not have to learn how to

1:10:11use code rabbit. So look, it looked it

1:10:14just resolved it addressed in commit

1:10:16this and you can go and look. Okay,

1:10:18here's where it was addressed. Um,

1:10:22so yeah, so that's the great thing. It's

1:10:23you do not have to learn how to use a

1:10:25tool like this. It's just it's just uh

1:10:28it's it's it's just it's the same exact

1:10:31interface of if you were dealing with a

1:10:32human reviewer on your PR, which

1:10:34[snorts] pretty much every developer,

1:10:36they know that workflow. they know how

1:10:37it works. Um, so and know one more thing

1:10:41I I'll provide because code rabbit has a

1:10:43lot of features and this is this is a

1:10:44feature that I also like. Um, I'm just

1:10:47going to res I'm going to respond to

1:10:49code rabbit in natural English. Um, so

1:10:52there's a bunch of things that you can

1:10:53do. Um, so I can respond in natural

1:10:56English and go and say, you know, let me

1:10:58provide you some more context and it'll

1:11:00go and re-review with that context. I

1:11:02can go and say I disagree with this and

1:11:04it will go and say it'll reconsider

1:11:07based off your prompt effectively. You

1:11:09could also do another thing uh where

1:11:11code rabbit has this thing called

1:11:12learnings where basically you provide

1:11:15some feedback uh on uh uh on a on a PR

1:11:22uh on an issue that it had brought up.

1:11:24Um and you can say oh this is actually

1:11:26not a thing. Here's some context of why.

1:11:30Um, so what it goes and does is it goes

1:11:32and takes that and it goes and say, "Oh,

1:11:33okay. You've actually taught me a new

1:11:35thing about your codebase. I'm going to

1:11:36file that away so I don't bring up this

1:11:39category of issue that is actually maybe

1:11:41not a problem." Again, um, so I'm just

1:11:43I'm going to say on this one, I'm just

1:11:44going to say I'm not um concerned about

1:11:48this issue. Uh this is this is a simple

1:11:52pet project and a a case that does not

1:11:57need to be uh

1:12:01handled.

1:12:04Okay. Um and then so you give it a

1:12:08minute, Code Rabbit will respond with

1:12:10the power of editing hopefully.

1:12:15Sorry, I've I'm giving you a humongous

1:12:17editing job. Um

1:12:20or you could just make this a three-hour

1:12:21video if you want.

1:12:22>> Yeah. No, it would be it's just like

1:12:24we're walking end to end from an idea

1:12:27designing architecture and code reviews.

1:12:30It's it's super fun and useful for

1:12:33anyone who wants to use these tools.

1:12:36>> Yeah. Yeah. Um cool. So, okay. So, Code

1:12:38Rabbit is has saw this issue um come up.

1:12:42It's got the little eyes emojis, so it's

1:12:44probably thinking about it. Um, we'll

1:12:46see how it responds.

1:12:49I don't know if GitHub will.

1:12:53Yeah, look at this. Understood. I'll

1:12:54make a note of that for future reviews,

1:12:56right? So, it has this like memory

1:12:59>> and it, you know, in the future if a

1:13:01similar category of issue comes up, it

1:13:03it it won't bring it up. Um, usually

1:13:05there's scenarios where it will

1:13:06actually, you know, there's times where

1:13:07where Code Rabbit provides feedback and

1:13:09it's like it's missing context because

1:13:11it's not inside of the, you know,

1:13:12context of the PR or it didn't, you

1:13:14know, build that context. It's missing

1:13:16some context. You could provide some

1:13:17context and it goes and stores that in

1:13:19its bank is basically. Um, so that's um

1:13:22that's basically it. I'm going to I

1:13:24think I'm going to merge this PR. I'm

1:13:25I'm happy with my review. Um, I fixed

1:13:28the only thing that I was like really

1:13:30concerned about. Um, so I'm going to

1:13:32merge this PR. You know, if I was in a

1:13:33company, I'd also wait for a human

1:13:35review. Um, but uh, you know, here we're

1:13:38just rocking it solo. So, look at that.

1:13:40Luna's Cat Cafe, the API is done. Um,

1:13:43>> beautiful.

1:13:44>> If we have the time, the last thing I

1:13:46want to do is I want to build a front

1:13:48end. And I'm just going to tell cursor

1:13:50build a front end for this and see how

1:13:54it how it goes. I think that would be

1:13:56fun.

1:13:56>> Cool.

1:13:56>> Yeah.

1:13:57>> All right. So,

1:13:58>> see it come to come to life.

1:14:01>> Yeah. I'm gonna So, I'm going to spin up

1:14:02my uh I'm going to spin up my back end.

1:14:05So, it has uh it has it available to it

1:14:09itself. I don't know if it's going to

1:14:10attempt to query it, but I'm going to

1:14:12I'm going to have it running. Um, all

1:14:14right. So, this is a Do I want to make a

1:14:18new No, I don't want to make a new a new

1:14:20thread because there's useful context

1:14:21about the endpoints inside of this

1:14:22thread. So, okay. So, we've got the API.

1:14:26Um,

1:14:27I'm going to make a new folder. Oops.

1:14:30I'm going to make a new folder called

1:14:35front end. Um, do you have a framework

1:14:37of choice? What are you usually uh doing

1:14:39your front end work in?

1:14:40>> Nex.js.

1:14:41>> Next.js. Cool. Let's use Next.

1:14:43>> Recently have been dabbling into

1:14:46Tanstack start.

1:14:48>> Cool. Yeah, we like we like tanstack.

1:14:51>> We like we like Nex.js. It's uh it's

1:14:54getting a little rusty. Oops. Not ls.

1:14:56I'm not in my terminal. Um, okay. So let

1:14:59me um let me run a command here. Okay.

1:15:01So let's build

1:15:04uh a front end for this API

1:15:09in the front end. I think I need to have

1:15:13a file in here

1:15:16for cursor. Um in

1:15:21yeah it won't recognize a empty

1:15:23directory. uh let's build a front end

1:15:25for this API and it should be a nextjs

1:15:29js project. Okay. So I would say so the

1:15:32homepage

1:15:34should be a list of cats. We should be

1:15:38able to to click into a cat and make a

1:15:44reservation

1:15:47uh to book some time with the cat. Uh

1:15:52there should be an adoption form page

1:15:57where you can submit

1:15:59an adoption

1:16:02request.

1:16:04Uh

1:16:07you might have been able to infer this

1:16:09from like these fields in the schema

1:16:12even to say all right maybe maybe

1:16:13there's an adoption application here

1:16:15because there is a field for it here.

1:16:17>> Yeah. All right. Let's let's do it.

1:16:19Let's do it. Let's um let's all we'll

1:16:21the only direction we'll do is we'll

1:16:23tell it to be an X.js project. I like

1:16:24that it's next.js because I know how to

1:16:26operate next.js as well.

1:16:28>> All right. Cool. Yeah. Yeah. Yeah. You

1:16:29you're you're probably totally right.

1:16:30So, let's build a front end. So, let's

1:16:32see what what is what is it what does it

1:16:34think that uh that that Luna's cat cafe

1:16:37should be?

1:16:37>> Um

1:16:39yeah, I I'm I'm such a I'm famously an

1:16:43over commenter in my code. Um you don't

1:16:45see that here because AI doesn't over

1:16:47comment. I'm an over commenter in my

1:16:49code and I'm like an over like providing

1:16:51context to prompter uh in

1:16:54>> I can tell I can tell I mean that that's

1:16:56how you you would get good quality

1:16:58results. Um I was going to ask if you

1:17:01ever use voice to kind of like whisper

1:17:05to voice to text but with the amount of

1:17:07references you create probably it's just

1:17:09easier for you to type because you're

1:17:11just like copy pasting highlighting

1:17:12putting it that in the cont that that

1:17:14would be impossible with with the voice

1:17:16thing. Yeah. Yeah. So, so I do not, but

1:17:19I have co-workers who do um and they do

1:17:22really like that because it's like uh

1:17:24one of my co-workers literally just

1:17:26brought up um this semi- recently where

1:17:28they were like, I want to be able to,

1:17:31you know, work on a thing when I'm like

1:17:32cooking and you know, I've got my

1:17:34headphones on that have a microphone.

1:17:36Like, I want to just be able to interact

1:17:38with with uh you know, my AI agent to

1:17:41like continue working on my codebase

1:17:43while I'm cooking dinner. And it's like

1:17:45I mean that's not a that's not a

1:17:49workflow that I myself uh think is will

1:17:53be additive to my life. Like I do have a

1:17:56pretty good separation. Like I'm I'm a

1:17:57I'm a I'm at like a a desk basically a

1:18:01desktop computer when I'm at work. It's

1:18:03technically a laptop, but I have like a

1:18:04desktop setup. I like having I'm at my

1:18:06desk and I'm working and I'm not at my

1:18:09desk and I'm not working. I think if I

1:18:11started introducing that workflow into

1:18:13my life, that would not be super great.

1:18:14But it's like I think there is something

1:18:16nice about being able to kick back on

1:18:18your chair and like lean back and just

1:18:21be like, let's work, you know, really

1:18:24relaxation mode. I think there is

1:18:26something to that. Um, so I don't use

1:18:29it, but I have co-workers that use it

1:18:30and co-workers who really like it. Yeah,

1:18:33I think more um also um besides the fact

1:18:36that you can multitask or I don't know

1:18:38speak while you're doing other stuff, it

1:18:41is also sometimes that you can speak

1:18:43faster than you can type. I mean you're

1:18:44you're pretty quick but usually humans

1:18:48can speak faster. So you can just get

1:18:50more um and then sometimes this uh tools

1:18:52like Visper, they can kind of fix what

1:18:56you're you're saying. you can just like

1:18:58um blur whatever you want to say and it

1:19:01just corrects what you wanted to

1:19:02actually say and it creates a better

1:19:04kind of sentence and prompt and I have

1:19:06seen that done but with but when you're

1:19:08providing references like you are and

1:19:11copy pasting and lines and ad signs and

1:19:13whatnot you can't I'm not sure if you

1:19:15can do that with the voice yeah I mean

1:19:17I'm willing to bet that like you can d

1:19:19you still direct it in the same way

1:19:21where you could say oh yeah make sure

1:19:22you reference you know how I'm using the

1:19:25repository function inside of my cat's

1:19:28repository, right? And and it could

1:19:30probably figure out where where you're

1:19:31going. Like, but yeah, I'm Yeah, I'm

1:19:33like I've I've I've always been this

1:19:35way. I've always been just like more

1:19:36information is better. And also too,

1:19:38this is this is a pro this is a pro tip

1:19:40for uh for those of you who are, you

1:19:42know, adopting AI and in your codebase.

1:19:45AI loves comments. Just like how humans

1:19:47love comments, AI absolutely loves

1:19:51comments. it you you're basically

1:19:53providing it inline contextual

1:19:55information when it goes and says oh I

1:19:57need to read this file if you add

1:19:59comments of like the wise and like why

1:20:01are you doing things it's like it can

1:20:02gain that context right as it's reading

1:20:04the code um and I've absolutely found

1:20:07that like my uh crazy over commented

1:20:11code bases AI works way better I can

1:20:14provide shorter prompts to it and get

1:20:17the result that I want because it can

1:20:18infer all of the context so it's like

1:20:20you know you're not only just writing

1:20:22comments for humans now like comments

1:20:23are really great for AI when you like

1:20:25hey I want to provide context you know

1:20:28inline every time an AI agent basically

1:20:29runs through this file it's like add a

1:20:31comment it's the same exact same exact

1:20:33way you would comment for humans um but

1:20:35it it it

1:20:36>> it it will you yourself uh will

1:20:39experience some benefit um in a way that

1:20:42you didn't um look at this it's it's

1:20:44totally it's totally building it um

1:20:48>> run build command to test everything is

1:20:50working there Yeah, I have a I have a a

1:20:54I have a side project which is vibe

1:20:56coded vibe coded and it's like I'm

1:20:58totally I'm totally constantly impressed

1:21:00by it. I'm just like how can it do this?

1:21:02Um so okay so the API is running on port

1:21:068080. So let's go and in another

1:21:08terminal here uh let's go uh cd frontend

1:21:12and mpm rundev.

1:21:17Oh it's just opening in cursor.

1:21:19>> Yeah. All right, cool. All right, so

1:21:22>> adoptions. There was that tab in of the

1:21:25>> Yeah. Yeah. Yeah. I didn't I did not

1:21:27tell it anything. Okay, so let's go to

1:21:28cats. Let's add a cat. All right, so

1:21:31we've got Luna.

1:21:33Um, she is a tuxedo

1:21:37cat.

1:21:38Let me see. Is it sending the Yeah, it's

1:21:40sending the requests here. Just did an

1:21:42options request. Uh, okay. So, let's

1:21:45just I'm just going to Oh, this is a

1:21:47future birth. Let's not do a future

1:21:48birthday. Let's say she was just born.

1:21:50Failed to create cat. What's going on

1:21:52here? Node.

1:21:56Failed to create cat. Let me see. Can we

1:21:57inspect in here? Uh, is that I've never

1:22:01used this. No, not components.

1:22:04>> Yeah. Okay, you inspect. So, let me

1:22:05look. So, network. Let's see what's

1:22:07happening.

1:22:10>> Uh,

1:22:13what? Next.js original. Oh, is this

1:22:15doing a server side request? Let me see.

1:22:18Failed to create cat. Give me some more

1:22:20information. All right. I'm going to I'm

1:22:22just going to say uh

1:22:25whenever I attempt to create a cat, it

1:22:30fails to create the cat. Oh, you know

1:22:33what? There might be a corores issue

1:22:34because it's running options requests.

1:22:36That's all I'm getting. Um let's fix

1:22:39this.

1:22:40Uh I Let me say Oh my god. Give it the

1:22:45context. Give me the context. Why is it

1:22:48not working?

1:22:50Is it

1:22:51okay? Oh, control L. Sorry. You can edit

1:22:55this out. I want to add to chat.

1:22:58All right. I think we are. Okay. So,

1:23:01this is an issue.

1:23:03>> This is an issue with the uh back end.

1:23:05So, the back end I don't have corores

1:23:07configured. So, uh it basically this is

1:23:10like a security thing for those of you

1:23:11who are unfamiliar with corors. I mean a

1:23:13lot of people are familiar with cros

1:23:14because a lot of people run into this

1:23:16problem.

1:23:16>> Um so I need to uh I need to open this

1:23:19up.

1:23:20>> So your front end is not hosted from the

1:23:22same kind of port that your back end is

1:23:24running. So browsers are like uh having

1:23:26that issue with the course. One thing

1:23:28about Nex.js that I find really really

1:23:30useful. They have this MCP tool called

1:23:32Next Dev Tools. Uh and in NexJ 16 you

1:23:36can just run that. it it has like an MCP

1:23:38server that gives your agent a window

1:23:40into the runtime server so it can read

1:23:43the runtime issues and provide more

1:23:45context. So sometimes you could just

1:23:47type in and say hey next dev tools tell

1:23:49me what's going on and you can see that

1:23:51there's like 29 issues there next is

1:23:53showing on the browser it actually can

1:23:56read all of those provided back to the

1:23:58LLM and then that just goes walks

1:24:00through those issues one by one and

1:24:02solves them.

1:24:04>> Cool. Okay, we verified it indeed was a

1:24:06cause issue because now I'm seeing the

1:24:08total cats, the total available for

1:24:10adoption. So, let's um I don't know.

1:24:12Let's let's mess around with this uh

1:24:14with this UI. Um let's make a new cat.

1:24:18Let's call him Fubar. Uh and he's just

1:24:22uh let's say he's an orange cat. Uh and

1:24:25he was born I don't know here. Okay. So,

1:24:30cool. So, we've got Fubar. Let's look at

1:24:32Fubar. Cool. He's available for

1:24:34adoption. No reservations yet. Let's

1:24:37book a reservation with Fubar. Oh, wait.

1:24:38We don't have any customers yet. Let's

1:24:40make a customer. Um, let's make a

1:24:42customer. Let's call Moran.

1:24:45He's loves cats. Um, let's go back to

1:24:50Fubar and let's book a reservation. So,

1:24:55Ron just gave us a call and he said,

1:24:56"Hey, I'd really like to meet Fubar uh

1:24:59like ASAP." So, it's 11:35 right now.

1:25:01So, he's going to be in here by noon.

1:25:03Um, so let's book a reservation. So Ron

1:25:06is going to Ron is going to meet with

1:25:08Fubar. Um, and uh, cool. Let's say they

1:25:11met. So Ron,

1:25:14okay, so no adoptions yet. So okay,

1:25:16let's say that Ron is really happy he

1:25:19met Fubar. So Ron wants to now adopt

1:25:22Fubar and he's going to submit an

1:25:24application. Um, and the staff really

1:25:28appreciated how uh, Ron treated Fubar.

1:25:32Uh, and they're going to approve the

1:25:34adoption. Unknown cat. Oh, a bug. Let's

1:25:37see. Let's see though if if the adoption

1:25:39actually happened happened. Uh, so Total

1:25:42Cats. Okay, look at this. Total Cats

1:25:44available for adoption too.

1:25:46>> So, so Ron has now adopted Fubar. So,

1:25:51view details.

1:25:53>> It doesn't have it doesn't have uh any

1:25:56any details about who the owner is, but

1:25:58that's okay. We don't need that. Um but

1:26:00so not notably though I know it did the

1:26:02thing because he now there's only uh two

1:26:06that are

1:26:06>> available. Yeah.

1:26:08>> Cool. It's great.

1:26:09>> Let's just close that. We're not worried

1:26:11about these issues.

1:26:12>> And uh

1:26:14>> Cool. There you go. My cat cafe is is is

1:26:17ready to open for business [laughter]

1:26:21>> live. All right. Beautiful. That was

1:26:22pretty cool. We built it end to end. I'm

1:26:25talking about how you're using AI in

1:26:27different um capacities, really tightly

1:26:30controlled small tasks with context, but

1:26:34also using the plan mode, using markdown

1:26:36files for the architecture to giving it

1:26:39more room um to go and build stuff based

1:26:42on the context you're providing on how

1:26:44you're doing things. We looked at your

1:26:47PRs and the code review with code rabbit

1:26:49that just goes in and kind of finds

1:26:51issues that you may not have noticed and

1:26:54then you can chat with code rabbit

1:26:56either provide commits that fixes those

1:26:58issues it picks up on those or you can

1:27:00just comment back and say don't worry

1:27:01about this. That was a very cool part of

1:27:04this whole workflow that you could just

1:27:06implement AI into whatever it is that

1:27:08you need to whatever extent you're using

1:27:09AI, you can just use that code review

1:27:12right now and then at the end building

1:27:14an XJS front end for it that just talks

1:27:17to this backend API endpoint.

1:27:19>> Yeah. Yeah. So we did it end to end. Um

1:27:22that would absolutely have not been

1:27:24possible in a one-hour uh meeting to

1:27:27build an entire API with a front end.

1:27:29Um, like so like there's definitely a

1:27:32it's it's like a big unlock for sure.

1:27:34Like obviously, you know, maybe I it's a

1:27:37little dubious of how production grade

1:27:38it was because I didn't review any of

1:27:40it, but it's like it works. It's like

1:27:41you got a proof of concept working. Um,

1:27:43and uh, yeah, so it's definitely uh, you

1:27:46know, we're in a new era for sure.

1:27:48>> All right, folks. So, that was it. If

1:27:50you have any questions, uh, hit me up

1:27:52down in the comments. If there are

1:27:53things that you do in your workflow, let

1:27:56us know and uh, we'd love to hear back

1:27:58from you. And hopefully we'll have

1:28:00Brandon more in future videos. And we'll

1:28:02see you in the next one.