Advanced AI Lessons using VSCode and Agents | AHK Hero Extract

0:00Join our Friday and Saturday Zooms for

0:01three hours of hands-on AHK teaching and

0:03support. Get unstuck and level up in

0:05real time. Um, connect with like-minded

0:07AHK users in our private chat, helping

0:09each other out and sparking ideas that

0:11make scripting way more fun.

0:15Now, uh, on the topic of coding with AI,

0:19right? So, oh, the okay,

0:25I have to figure that that one out. But

0:27in the topic of coding with AI and I

0:29actually mentioned this to Joe. Um

0:35I

0:37have discovered that having kind of like

0:40a pattern when you're coding and using

0:42the right

0:44uh model for that

0:47makes a ton of sense. It's it's it has

0:49been helping me. So the pattern that I'm

0:52doing is this

0:54for now. I'm asking questions,

0:58then I'm planning what I'm going to do,

1:02then I'm editing the file. All right?

1:05So, and that's basically what I do

1:08myself whenever I'm going to code a

1:11function. I first of all

1:14uh think about it, talk about it, like

1:17plan what the function should do, what

1:18it's going to do. I just think about it

1:20first

1:22what I want to pass to it what I'm going

1:23to return

1:25then I code right

1:29where AI is really great is at this

1:32questioning assumptions part because it

1:35has more knowledge or more context than

1:38I do at a given time so what I do is I

1:42start asking questions okay so I'm going

1:44to do this this and that what am I going

1:46to what is the best way it replies with

1:49that and that is already in the context

1:51then I say okay let's plan an approach

1:54when it comes back with it now here's

1:57the part for asking I don't need a very

2:01powerful model I just do something like

2:04um you know GPT5 mini or something that

2:07is good at replying and it's just giving

2:10me idea but when I go to the planning

2:13section okay let's create a plan on how

2:15to implement that I switch to a thinking

2:17model and one of the best ones that I

2:19have available. Right? It doesn't matter

2:22which, whatever you prefer. Claude. So

2:24now remember Claude has haiku which is

2:27the client kind of like the the fast

2:30model.

2:32Then you have the middle one and then

2:34you have the one that is kind of like

2:36oppus I think which is the the mo the

2:38thinking model. Well for asking

2:41questions I will grab the middle one.

2:44Okay. Just because it's good at just

2:46replying and stuff.

2:48Then the next one when I'm going to plan

2:50ahead I'm going to say okay let's think

2:52about this problem give me the potential

2:55steps that I'm going to take then I

2:57switch to the thinking model so that it

2:59grabs all the context it does everything

3:01and then once that is ready then I

3:04switch to the fast model this one

3:07doesn't have to think I don't want it

3:09thinking I want it just doing the

3:11following the direction that I just gave

3:14it and that's something

3:17Just going in that route right that has

3:20helped me because if you use a thinking

3:23model for coding what is going to happen

3:25is every time it's coding it's going to

3:28rethink everything and say let me change

3:30this as well and then it changes

3:32everything away right but if you use a

3:34model that is not thinking it just

3:37follows the instructions change line one

3:39change line five change line 10 and it

3:42doesn't modify your file as much so this

3:46might have obvious for all programmers

3:48out there,

3:50but as I'm, you know, more focused on

3:53other things and I usually do the coding

3:54myself, I've never actually thought

3:56about using the tool that way. And I and

3:58I have mentioned this before. Now, let

4:00me show you a real life example of that.

4:05>> I just for a second and say there's

4:07absolutely no difference in what you

4:08just said as to swapping out people with

4:12the different AI models, right?

4:14>> 100%. some people, you know, to go to

4:17for ideas, you know, and then there's

4:18some the last time when you want to get

4:20something done, you're like, "No, no,

4:22no, no. I just

4:22>> I

4:24got to do that with Exactly." Now, what

4:28I have discovered is

4:30the less specific I am,

4:35the higher syn thinking model I use.

4:39Okay? So if I give you a lot of very

4:42specific things, I would use a model

4:44that doesn't think that much because I

4:46already did the thinking.

4:46>> You're given the context,

4:48>> right? Because I'm giving it everything.

4:50But if I am very very vague,

4:55then I would not do that. Let me show

4:57you here. So this is uh part of a tool

5:04that I just created that is for reading

5:07a bunch of log files and generating a

5:10report out of it in the end. So that's

5:12it. It just looks for log files. It

5:14counts them. This part I coded myself. I

5:18have been doing this so long that I know

5:20what I'm doing. I just did this. I just

5:22did that. Uh I selected the dates. I

5:26created the G the the uh regular

5:29expression. All good. I already created

5:33my my my

5:35stuff up to this point. At this point

5:37115,

5:39I was so tired because I this was at the

5:42end of the day that I was like, "Yeah,

5:45now I have to grab all this object that

5:48I have and now I have to flatten it to a

5:52CSV." and I know what I can do for it,

5:54but let's have AI do that for me.

5:59First thing, I need to flatten the data

6:02for a CSV file. Is that possible? Show

6:04me five example rows. This is a very

6:09vague statement. So, what I did when you

6:12look at the end of this, I use Gemini 3

6:16Pro. It is a very highinking model. I am

6:20in asking mode. And let me talk about

6:22that a little bit. In VS Code, you can

6:25switch the mode what you're working

6:29with, asking, editing,

6:32being an agent, or planning.

6:37This is VS Code stuff. So

6:40when you're asking,

6:42the AI doesn't have access to any tools

6:45and it cannot modify the file.

6:47So it can reply back to you but it will

6:51not modify the file and it has access.

6:53It doesn't have access to almost any

6:55tool exert I think except for the online

6:58search I guess because it's for asking

7:00questions. I made that question in

7:03asking mode because I don't want it

7:05changing anything. I just want it

7:07thinking about it and figuring out what

7:09we're going to do. He came up with an

7:11idea. Okay, this is what it look what it

7:14would look like in the end. And I was

7:16like, "Yeah, that looks cor it suggested

7:19some code changes, but it cannot change

7:21the code." So, it just suggested it. I

7:23said, "Yeah, that's good. I like the

7:25general idea, but let's do this. I want

7:27to do this." I I I explained something

7:31else that I wanted to do. Still, when I

7:35look at the bottom of this,

7:37it suggested a bunch of stuff when it I

7:40am still using Gemini Pro. I'm still in

7:44planning mode. All right, I'm just

7:46asking questions, reviewing what I'm

7:49going to do. So, it said, "Yeah, that's

7:52a that's a good approach. This is a two

7:55pass algorithm because I I I told it,

7:57hey, I want to do it in two passes. One

7:59of them, the first pass is to figure out

8:01what the headers are going to look like

8:03in the end, and then we're going to

8:04flatten the data into that." Something

8:07like that. It said, "Okay, that's good.

8:09Let's do this." Blah, blah, blah. It

8:11gave me some suggestion code. And then I

8:14look at the code and I say uh yeah that

8:17is not needed because it was trying to

8:20sort the statuses. But the funny thing

8:22when you look at the code is that it

8:24created a map. So the first pass is

8:28reading the stuff. I think it was it was

8:33it creates a map right here and then

8:37later down the line it tries to sort the

8:39thing. And I'm like that's dumb. maps

8:44are already sorted. So that's the cool

8:47thing about looking at what AI is

8:49suggesting because you can come up with

8:51these kind of things. And then the next

8:53one is uh there are some checks that

8:55needs that are redundant. I don't care.

8:57And instead of using file append like

9:00file delete and then file append just

9:02use file open with W which does that.

9:04Okay. So I I I'm guiding what the end

9:07code is going to be from the beginning.

9:10It tries to do some validation. it comes

9:12up with, hey, you know what? You still

9:15have to sort the headers because uh uh

9:18uh auto hotkey v2 are insertion ordered.

9:22I'm like, no, you're you're wrong. I'm

9:24doing all of this before even writing

9:27before even telling it to write a line

9:29of code because that way when I tell it

9:32to write code, it's just going to follow

9:34the instructions. I say, "Hey, you're

9:36wrong. Follow my instructions. remove

9:39step two and just keep the step simple.

9:43It put it in as I mentioned. Perfect.

9:47Now, at this point, I tell it build a

9:50to-do list and implement step one. Just

9:53one step. Don't implement the whole

9:56thing. Just just one thing. That one

9:59when you look at the model now

10:02is a mini

10:05it is a mini uh model. is not the

10:09thinking model that I was using for

10:10planning. So for asking question and

10:13planning probably use a model that has a

10:16lot of context that can think ahead that

10:18can plan that can think guide all your

10:21stuff and always do it without having

10:24access to editing the file.

10:27After you're good with what you what you

10:30were what was suggested to you,

10:34then you use you I switched. Now in VS

10:39Code, you have either edit or agent. As

10:43you can see, agent has access to tools

10:46down here, but edit doesn't. So edit

10:52allows you to t kind of like target your

10:55edits into one file. That's it.

10:59If you use agent, what happens is that

11:02it will try to determine which files it

11:05needs to modify. It will do terminal

11:08calls to find out like on the terminal

11:11like the the the the terminal down here,

11:13it would try to create commands to

11:15search for files and what files it's

11:17going to modify and stuff. So it might

11:19go a little bit crazy. So no, I actually

11:22made it edit because I know that it's

11:24just the edits are going to be just in

11:26this file. It went ahead and implemented

11:29number one. It looked good. I just asked

11:32it a few things. Why do you need a

11:34separate that one map can do that? It

11:38explained why and I said no, you're

11:40right. Keep it as is. Uh the map doesn't

11:43preserve insertion code. So keep both

11:46one for the sort of this. So keep it

11:47like that. And then from there we

11:50continue. Remember for those kind of

11:52things I I make it in asking mode.

11:56Whenever I don't want it to modify the

11:58file. I'm just asking questions I switch

12:01to asking mode. And you notice that I

12:04already have hot keys for everyone.

12:07So I go by steps. First I'm going to

12:09ask. Second I'm going to plan. Third I'm

12:12going to do an agent. Fourth is the most

12:14kind of like targeted thing. So I

12:17already have hotkeys for them to just

12:19switch between them very quickly.

12:22And then I just say let's continue on to

12:24the next step. I verify the code that it

12:26gave me gave me. Very good. Generate the

12:29next step. Very good. Okay. Use format

12:32time for something else. That's it. Just

12:35by doing that at the end I spent like I

12:40don't know 15 minutes in that process.

12:42So

12:44and in the end the code works as I

12:47expected. I already generated my GUI. I

12:51already had it so that the the time

12:53selection works. So I just select the

12:55selection. I just select a minimum

12:58version. I select it reads through all

13:02the files and then

13:05once it finishes up it just creates an

13:08Excel report. well, a CSV file and opens

13:11it with Excel because that's my default.

13:13Right now it's very slow because I'm

13:15sharing my screen, but that usually

13:16happens in like less than a second, less

13:19than 5 seconds, I guess. Um, and once it

13:22finishes doing that,

13:25I have the data in the format that I

13:27want. So, I have found this approach to

13:32be better. I'm not using it all the

13:34time. It's not like I'm going to

13:36implement the whole program because as

13:38you noticed just for one little thing

13:42the only thing I needed to do is flatten

13:44the CSV right flatten the data into a

13:47CSV just for that look at all the steps

13:50that I had. I had like five different

13:53things and look at all the little

13:55details that I had to fix.

13:58In a real full program, you will have

14:02hundreds and hundreds of steps and

14:04hundreds and hundreds of little details

14:07that for each of them you have to go

14:09step by step.

14:12So I would definitely not use it for hey

14:16let's create a program and just go for

14:18it. I would say hey I would plan what

14:21the programs are. Look at this. This it

14:23this created exactly what I needed. And

14:25basically now I just select this row and

14:29say, hey, let me conditional format that

14:33on on if the number is greater than 10.

14:38I'm just looking for failure rates that

14:40are too high. See what I mean?

14:43Everything below 10, good failure rates.

14:46Okay, let's look at that computer and

14:48see what's going on. A lot of single

14:50instance force in there. Why? what's

14:53going on. You see, that allows me to do

14:55this kind of thing really quickly. And

14:57again, going back now to the topic of

15:00using AI to code

15:02and the the video that I sent recently

15:05and Joe, what you were just probably

15:07want to talk about, which probably we

15:09have a good segue, is planning ahead

15:13removes a lot of the problems that you

15:16will have later. So if you're using AI,

15:19if you can plan the stuff without

15:22modifying the file first, right? And

15:25once you have the plan that looks good

15:27for you, then and you select a a a lower

15:32model to edit, then the lower model is

15:36not going to change your code in weird

15:38ways. It probably will just follow the

15:40instructions of the map,

15:43which in turn, at least for the test

15:48that I did last night, because again, I

15:50haven't been using AI to code like this.

15:52I'm I'm testing it from time to time

15:54because I didn't like it. It was just

15:55doing weird stuff. I I just started to

15:58find out, okay, this process seems to be

16:01a better approach and it seems to be a

16:04nice tool to have for creating stuff

16:06that I didn't want to create. I didn't

16:08want to at 11:00 p.m. in the night think

16:10about how to flatten a freaking weird

16:12object that I created. Let AI do that.

16:15Figure it out. And now that you figure

16:16it out, write the code and I would just

16:18review it. You know, that's a good

16:20simple way to do that.

16:23No, I think it's great that you have

16:24that control to be able to say you don't

16:26have read and write access. I mean, you

16:28can read, but yeah, but you can't edit

16:30like that. That I know from you and Nick

16:32just talking like that's a game changer

16:34in itself, right? of like just

16:37>> just

16:37>> I can I can say Isaas I am doing the

16:40same as you are doing.

16:43>> Okay, that's awesome. And actually I

16:45noticed in the S code

16:47>> you can create your custom agents.

16:50>> Yes. Yes.

16:51>> So those that I saw there the ask uh

16:54plan whatever you can create your own

16:56mode and you can specify exactly which

17:00tools you have access to. So it g that's

17:03what I'm working on right now like

17:05whenever I have uh free time I just go

17:08into VS code I talk to it because I have

17:11been starting to use the hotkey to just

17:14dictate and I would just dictate look

17:16the agent has to do this this this this

17:19and that add it into the file and it

17:21just goes ahead and adds it you know so

17:23so

17:25it has been a an interesting experience

17:27for me now that I'm doing that and I

17:29guess um by the way let me show you

17:32right now. One of the things that I

17:34discovered today is that AI, at least

17:39within VS Code, I don't know if AI in

17:41general can do that. So, if you go to to

17:45chatgpt.com or claude.com,

17:48they might not have access to this

17:50GitHub repository

17:52because the repository

17:54belongs to Microsoft and they they

17:57probably limit how AI access them,

17:59right? But within VS Code, which is also

18:02from Microsoft, I was able to say, hey,

18:06can you access this? It said yes. And I

18:09actually confirmed, go ahead and and and

18:11bring me some lines of code from that.

18:14Right? So, and it actually returned back

18:17lines from the documentation using

18:19Python, right? And I was like, okay, so

18:22you can access it. Look in that URL, we

18:25have the documentation for alpha version

18:28two and version one.

18:31select the default branch and every time

18:34you're going to be replying to me, try

18:37to read from that documentation. So, we

18:39have been having this issue Joe that we

18:41haven't been able to tell tell it to

18:43read the documentation, right? But now

18:45that I figure out that it has access, at

18:48least through VS Code, it has access to

18:50that repository, that might change. And

18:52I'm trying to force it to every time it

18:54replies, go ahead and get the relevant

18:57files, extract the information that it

19:00needs, and then reply. So

19:03>> to make to make it clear uh Isaas

19:06does uh does AI really has access to it

19:13or is uh AI feed with it? So I mean is

19:18does AI have the uh information about

19:22now

19:24or since a months?

19:26>> No. No. So that is that is a great

19:28question. So

19:30>> the last the last months I I I mostly I

19:34I recognize that AI doesn't know about

19:37the last months.

19:39>> Yeah, I can't. No, but if you use the

19:42fetch tool, so you can see here I'm

19:46using the fetch tool.

19:47>> Okay. Then then it is okay.

19:50>> That will go to the website, fetch the

19:53data, right?

19:55>> Okay.

19:56and then using

19:58>> page that data and then

20:01I told it give me the lines. So what it

20:04did it goes ahead and fetches the

20:06information. All right. So it goes and

20:09tries to download the website.

20:11>> Okay.

20:12>> Right.

20:13>> And then it uses Python on the file that

20:17downloaded.

20:18>> Okay.

20:19>> And then it gave me the lines that it

20:22got from the internet. So when I looked

20:25at it here for example, this is what I

20:28was trying to figure out. Look at that

20:30version 2.1 alpha.

20:33While for the other file, the version

20:35was different here 2.09.

20:39So it was able to fetch the current

20:43information on that URL, extract the

20:46information, and when I looked at the

20:47version number, they were both

20:49different. So I knew that it was able to

20:53switch branches and that's basically

20:56what I'm now having in my auto hotkey

20:59instructions. Hey, this is where the

21:02documentation will leave for you and for

21:05all my agents. So whenever I switch

21:07agents to thinking or planning,

21:11I'm having this uh requirement there.

21:15It's also because of yesterday uh I

21:18asked for uh I asked Lo to give me the

21:23uh differences between Gemini 5 and

21:27three and uh sh uh GPT 5.2 two.

21:34>> Okay.

21:35>> And clothe arms with me that the highest

21:39uh

21:41of was four

21:44>> and uh of Gemini was 1.5. So that's

21:48that's what I'm thinking. It's it's it's

21:50it's

21:53>> you didn't enable the search

21:55functionality.

21:56>> I see. I see. I see. Okay. I would I

21:58would say n it sounds like you also

22:00weren't using a thinking model because a

22:02thinking model

22:03>> probably would have said hey maybe there

22:05is a newer version let me go look

22:06>> okay okay

22:07>> right

22:08>> the model so so that's why I'm

22:10discussing this if you're not using the

22:12correct model and if you're not using

22:14the correct tools with that model you're

22:17going to get funny answers so

22:20is is is a good idea to just and again

22:23all of the all of the versions that I've

22:26all the chats Oh, sorry. The the LLMs

22:29that I've tested have usually three

22:32levels. They have the mini that is the

22:34fast thing that doesn't think. It just

22:37replies. That one we should use it for

22:41actions. Do this, do that. All right.

22:44There's a middle point in which you have

22:47GPT 5.2. There's the the fast or mini,

22:53the GPT 5.2, 2 and then there's the

22:55thinking mode of GPT.5.2 two for example

22:59and when you have the three then you

23:01just pick depending on what you're

23:04trying to do

23:04>> and to your point Isaas when I think it

23:07was five from Chad PD came out they did

23:10away with what you just described they

23:12said

23:14>> and then everyone they're like no no we

23:16know what we're doing and then everyone

23:17apparently said I want access to that

23:19older one and control and they messed up

23:22and said you guys are right we'll we'll

23:24back it off

23:25>> we we'll allow you to do that because

23:27the problem is I like I I want the finer

23:30finer control

23:33and I at least for now the way how I

23:37code is I think big. So I use a planning

23:40model then I ask questions I use a

23:44smaller model and then I edit. I use the

23:48smaller thing. So I go from big to small

23:51and I I'm changing my model as I go. You

23:54throw a wrinkle in what you just

23:56described though Isa as right.

23:58>> My understanding is the new Gemini flash

24:00model is smaller than the full Gemini

24:04one but it's actually better at

24:06programming which is really

24:07>> that is correct. That is correct. So my

24:10problem is within VS Code and that's why

24:13I was replying to uh Connie like I did

24:16uh earlier is within VS Code I've

24:19noticed that whenever I try to access

24:21Google models it has a lot of failures

24:25but if I go to their website everything

24:26is fine that's when I realized okay so

24:29probably VS code Microsoft wants you to

24:31use their own models you know like so

24:34they they would make it a little bit

24:35tricky to use other models or they don't

24:38have enough access or uh they have some

24:40limits something but while I'm trying to

24:43use VS code I haven't been able to uh be

24:48content of using Google uh uh Google

24:50Gemini flash or thinking

24:54>> but I don't say that they're bad every

24:57time I've used Gemini and actually in

24:59that example that I just showed you I

25:01was using Gemini for planning it was

25:03really good at planning like really good

25:05at at planning so u I I definitely think

25:09it is a great model. It's just that

25:11through VS Code, I I don't have good

25:12access to it right now.

25:15>> Okay. All right. Let's uh stop the

25:17recording here real quick.

25:20>> I'm going to stand up for a second. Be

25:22right back.

25:24Ask questions, plan, and stuff. You

25:27don't have to do anything. So, you just

25:30open VS Code.

25:33You install the GitHub copilot chat

25:36extension

25:38and it comes with those predefined.

25:41As soon as you have your GitHub copilot,

25:44it will ask you to log in to either

25:47GitHub or your Microsoft account. If you

25:52don't have a GitHub,

25:54well, you probably have a Microsoft

25:57account because you're logging into your

25:58computer. It would be your login

26:01informations for your computer, right?

26:03So, I usually use my Hotmail account to

26:08log into Microsoft.

26:10If you have a local account and you

26:12don't have a Microsoft account, which is

26:14strange.

26:15>> I have I have a GitHub. I have a

26:17Microsoft.

26:18>> So, I usually set it up with GitHub, but

26:22you can set up with either one. Right.

26:24Second of all, they have a plan. They

26:28have a free plan. The difference there

26:31if you're going to use the free plan is

26:34that of course you have a lot of limits.

26:36How many times a day you can do stuff

26:39blah blah blah.

26:41If you use a paid plan

26:44you will have better limits. I am paying

26:47my plan is like $10 a month. is not uh

26:51too expensive and in my case it makes

26:55sense because I am coding 8 hours a day

27:00in VS Code or more so it makes sense for

27:03me to pay for it. If you don't want to

27:06pay, just know that you have a free

27:07version available.

27:09Some sort of limits apply.

27:14But

27:16once you install the GitHub Copilot,

27:20um you have a tab and usually the

27:23default hotkey is Ctrl Alt I. I keep the

27:26defaults for convenience. You can change

27:29that to whatever you want. And because

27:32the

27:34hotkey was Ctrl Alt I, I decided that

27:39all the hot keys for my agents.

27:43Again, once you install it, those four

27:45agents are going to be there by default.

27:48You can configure custom agents, which

27:51is what I'm in the process of

27:52configuring the agents to do the stuff

27:55the way I wanted to do them. But that

27:57takes time, so I'm I'm doing it slowly.

28:00But you can set hot keys for each of the

28:03agents. And I decided if Ctrl Alt I is

28:07going to bring the chat then Ctrl Alt

28:10number pads is going to switch between

28:12the agents and Ctrl period which is what

28:15I set up for this one here is to pick

28:19the model. So I just do Ctrl Alt 1 2. So

28:25uh let me see three four five. Right. So

28:28between so between those I can switch

28:30the agent modes just with my keyboard

28:33and if I want to change the model

28:35control period and it just gives me the

28:37menu and I pick one. So it makes

28:41>> how did you open the window on the

28:42right?

28:44>> Ctrl Alt I

28:48>> and that's the default hotkey. All

28:50right.

28:51>> How could you do that by the menu?

28:54>> Sorry. How could you open the window on

28:58the right using the menu?

29:01>> Oh yeah. So you just press F1 and just

29:04type chat and there's a lot of options

29:08that it gives you for chat either the

29:11chat settings uh switch to the next

29:13agent uh to open I think you have for

29:18opening new. So here open the chat open

29:21the ask open the edit open the plan.

29:25So, it depends on the ones that you want

29:27to use and you can set up the hotkeys by

29:30just clicking on the configure key

29:33banding. All right.

29:34>> Can you can you do it using the file

29:37edit selection menu?

29:40>> Give me just a second. Sorry. But yes,

29:42you can.

29:45We can run into this next week or I'm

29:47sorry, maybe tomorrow. This is we said a

29:50couple times that we would handle this,

29:52you know, later.

29:53>> I guess you didn't see that. All right,

29:55thanks everyone. Have a good day. Bye%

29:58off all our HK goodies, courses,

30:00scripts, and custom work so you save

30:02cash while building your toolkit like a

30:03boss. Tap into over 466 plus hours of

30:07AHK video gold, including fresh clips

30:09like this for anytime learning that fits

30:12your