How I Actually Code with AI as a Senior Software Engineer

0:00Hey y'all. Welcome back to the lab. In

0:01this video, we're going to be discussing

0:02how I actually code with AI as a senior

0:05software engineer. So, it's been a

0:07couple months since I posted stop vibe

0:09coding and I've been using AI daily at

0:11work and in my side projects. I wanted

0:14to give an update on how I'm actually

0:15using AI in my workflows. First, what am

0:18I using AI for? So, I'm using AI to code

0:20daily, not in every task, but in lots of

0:23them. Some areas where I think AI is a

0:25great tool for coding. First is

0:26investigations and research around

0:28coding topics. Second is in actual

0:30coding faster prototyping and

0:32implementation especially for

0:33boilerplate heavy code as we'll talk

0:35about later and finally for reviews

0:37quick first passes to turn a first draft

0:39into a second draft investigations. So

0:41AI has largely replaced Google for me

0:44for common code related searches. So

0:46examples of this might be give an

0:47example of a try catch in language X.

0:49Let's look up an error with text Y. How

0:51to do an exhaustiveness check in

0:53Typescript things like that. And this is

0:55something that I predicted a year ago in

0:56how software engineers actually use AI

0:58to improve productivity and has

1:00generally remained true. AI is simply

1:02faster at information retrieval and can

1:04mold the answers to better fit your

1:06exact query which beats out sifting

1:08through several adinfested SEO

1:10tangentially related articles to try and

1:12craft an answer for yourself. Of course,

1:14AI still gets things wrong and is

1:17typically better at more surface level

1:18queries, but the speed on first

1:20retrieval is generally worth the

1:22trade-off as the harder queries that you

1:23know AI does get wrong would require

1:25additional research anyway and so it's

1:27already going to take you extra time.

1:28For those, I typically fall back to

1:30reading the docs because if the AI is

1:32getting it wrong, then it's likely many

1:33of the initial articles are wrong as

1:35well. And so, you kind of got to go to

1:36whatever that's called like a first

1:38resource um to get the actual answer.

1:40So, some tools I use for this is mainly

1:42Claude via the the chat. And as you can

1:44see, I have it bookmarked here because I

1:45use Claude so much. And then I'll fall

1:47back to chat GBT and Gemini depending on

1:49what's available. Um, for instance, at

1:51work, we don't have a cloud subscription

1:53for um the chat part of it. And so, um,

1:55we use GPT and Gemini instead and

1:58they're both decent. Um, although I do

2:00like cloud better. Now, for coding. So,

2:01I use AI regularly to help code

2:03features. It's still not great at

2:05building large features in my opinion,

2:07but it is getting quite good at well

2:08scoped tasks and parts of features where

2:10examples exist for it to follow. And I

2:12think these well scoped tasks and places

2:14where it has examples are really

2:16critical for it doing a good job. And so

2:17that's really something to look out for.

2:19And in general, my process still follows

2:20my vibe engineering cycle, which you can

2:22read more about here. Uh the first is

2:24don't outsource the plan or thinking to

2:26AI as it's really not great at creating

2:28and maintaining direction or vision long

2:30term. And so that's going to be

2:31something that you still need to do. you

2:33still need to have an overarching goal.

2:34Um, what are we trying to achieve? How

2:36are we going to do it? What are the big

2:37milestones? Because AI just like cannot

2:39do that. What you should be doing is

2:40outsourcing specific, welldefined coding

2:44tasks to AI where it can be a big speed

2:46enhancer over manually typing this stuff

2:48out yourself, especially for things that

2:50are just tedious, um, things with a lot

2:51of boilerplate. And so, this is often

2:53going to mean that you need to build out

2:55like a first version of this, a proof of

2:58concept or something, um, initially so

3:00it has something to work off of. But if

3:02you've got 10 things all following the

3:03same pattern and you do one, AI is

3:05pretty good at doing um one to 10. And

3:08then finally, uh you should checkpoint

3:09your work regularly to avoid AI trashing

3:11all the progress you made. I see a lot

3:13of people they get like super deep into

3:14a feature, it's almost done, and then

3:16they ask it to like fix one thing and

3:17the AI kind of goes on like a death

3:19spiral and trashes everything they've

3:21done because it's trying to fix

3:22something else. And again, it's kind of

3:24lost that long-term direction and vision

3:27of what it was trying to do in the first

3:29place. And so lots of checkpoints are

3:31like really critical for not having to

3:32backtrack. Um if you play video games,

3:34you probably also incessantly save,

3:37although new video games are better at

3:38saving for you. And so you should really

3:40do that with AI here um as well. And so

3:42here's the kind of cycle that I use from

3:43like a high level. You know, I plan my

3:45project myself. I'm using AI for

3:47research and reviewing the project and

3:48you know, iterating on it. But um

3:50basically the the whole plan for a large

3:52feature or project I do. And then in

3:54each milestone and usually really a sub

3:56milestone, so a sub part of that

3:58milestone, I'll do a tight loop with AI.

4:00So um I'll give it a specific prompt. Um

4:02this will be like, hey, let's build out

4:03this part. Let's build like a unit test.

4:05Let's make this thing do this one thing.

4:06And so if it's very small and and

4:08isolated and pretty simple, I'll do just

4:11directly in the chat um because it's not

4:13too much to like put that in again if it

4:15if it fails. But if this is something

4:16that requires more context, like hey,

4:18reference this file over here and this

4:19file over here, then I will use a

4:20markdown file. And the reason I like a

4:22markdown file is if the AI goes off the

4:24rails, which it often will, um, it's

4:26easy to just be like, let's revert those

4:28changes and then I will make the prompt

4:30better to give you the context that you

4:32were missing and make you do it again. I

4:33always have it output its plan to a

4:35markdown file and then I'll review its

4:37plan and then make changes as necessary

4:39to the prompt. Um, so that it better

4:40understands what I'm asking it to do.

4:42And then I'll auto accept all edits. I

4:44think this is really critical because if

4:45you're trying to manually accept edits,

4:47you don't even see like the full thing

4:49that it's trying to do and it will also

4:50need to iterate before it gets to a full

4:52thing. It's kind of like jumping in to

4:54an engineer's feature and they're not

4:56ready to put it up for PR and like

4:57you're just going to get in the way of

4:59them getting to something reasonable

5:00that you should look at. And then at

5:02each juncture after it's done all the

5:03edits, I then review the code because

5:05that's a complete unit of something it

5:07thinks is is ready. And I'll trash it or

5:09change it. um trash it if it's like

5:11totally off and I need to actually

5:12probably change the prompt or go about

5:14it a different way, change it if it's

5:16close but it's just missing a few

5:17things. And then if it's made decent

5:18progress but still has errors, then I'll

5:20actually accept it and basically put in

5:22a temporary commit so that I can iterate

5:25on just the the parts it needs to change

5:27without getting rid of all the progress

5:28it's made. And that's kind of part of

5:29the checkpointing cycle that I use with

5:31AI. And then finally, I'll iterate to

5:33the next bit of the feature. Um, often

5:34it's going to be multiple passes with AI

5:36because it's like, hey, do this specific

5:37thing and then do this and then do this

5:39to actually get the full milestone out

5:41of the way. Um, but this way I think

5:43it's like heavily guided and so I am

5:46actively involved in the whole process.

5:47But I do get the benefits of like, hey,

5:49it's actually going and typing out all

5:50the code and doing the tedious work of

5:52that. Um, which I think does lead to

5:54some good compromises of like speed

5:56benefits, but also like not having to

5:58trash so much code because it just went

5:59off in the wrong direction. And I do

6:01want to call out that this is pretty

6:02similar to like my own coding process

6:04and I think how most engineers do um

6:06their coding workflows as well where

6:07like you start off by making a plan and

6:09then you're going to iterate towards the

6:10plan in very small batches. I'm a fan of

6:12atomic commits. So this idea of like

6:14we're going to commit progress and

6:15milestones that are part of a larger

6:17milestone we're doing regularly and

6:19trying to push that into main. I I like

6:20that. And then finally reviewing the

6:23code at each juncture and making changes

6:24is necessary because your first patch

6:26just isn't going to be right. as you as

6:28you move on, you're going to learn new

6:29things about the code and what you're

6:30trying to do and um edge cases,

6:32assumptions that you had. And so, you're

6:33going to have to go back and review

6:34anyway. It's just this time, you know,

6:36the AI is also in the loop here. I'm

6:38currently sticking with Agentic

6:40Workflows for coding. I personally don't

6:42like autocomplete features, whether it's

6:44for code or like in my text messages or

6:47like in my emails or like Google Docs.

6:48Like, I do not like autocomplete

6:50features because it breaks my train of

6:52thought. Um, which slows me down and I

6:54think just leads to worse quality

6:55overall. I do like agentic workflows

6:58because they can be called when I want

7:00them. So, I access them when I want

7:01them, they're gone when I don't, and

7:02they work in the background if it's my

7:04workflow better. I can go send it to

7:06like, hey, go do this thing, and then I

7:07can go plan something else or look into

7:09another bug I'm trying to fix or

7:10something, which I think is actually how

7:12I'm getting a little bit more of the

7:14speed benefits. And then I'm primarily

7:16using cloud code for this. I am open to

7:18experimenting with other tools like I

7:19want to try codeex at some point. Um,

7:21but cla code has just been the one

7:22that's been working well for me. I

7:24previously used Rode and Klein, but I

7:25just found that Cloud Code was smoother.

7:27It had a better integration and

7:28experience for me. And it also had

7:30better billing. I'm using Max right now,

7:32which has been really awesome. Although,

7:34you really got to check these like

7:35subscription things because sometimes

7:36you're just overpaying. Um, you can use

7:38like CC usage to check like how much

7:40you've actually used and see if it's

7:42actually worth the the plan that you're

7:44bought into. Reviews. So, I've started

7:46using AI as a first pass reviewer. And I

7:48got to say it's surprisingly good in the

7:50sense that it regularly catches things

7:52or has ideas that are useful. And I I

7:55think this is less up to how good the AI

7:56is and more just like having a second

7:58pair of eyes on a work. I have my own

8:00practices for reviewing my own code and

8:02my own writing to try and catch more of

8:03this stuff. So, for example, I like to

8:05review my code outside of the editor and

8:07in um GitHub or where wherever we're

8:10doing code reviews um because it helps

8:12me get out of like the code how I've

8:13seen it and kind of see it from a

8:15different perspective and that's often

8:16helps me catch more things. Similar with

8:18my writing, I don't review it in my

8:20editor. I'll usually review it in a

8:22different view whether it's like on my

8:23website or um in a different kind of

8:25markdown viewer just because it like

8:27looks different and so you'll actually

8:28catch more things that way. And I think

8:31the AI is pretty good. like it's not the

8:32best reviewer in the world, but it's

8:33pretty good for a similar reason. It's

8:34just like another perspective on the

8:36thing that you've written and so it's

8:38just going to catch things that like you

8:39didn't see or didn't even think about.

8:41And so my process for this is pretty

8:42simple. I'm just reviewing the code

8:44myself and I iterate like I always have

8:46done. But then I ask the AI to review

8:47the code itself and then provide me

8:49three things that could be improved and

8:51then I iterate on those. And some of

8:53these are useful. Usually like one of

8:54these is like, oh that's that's a good

8:56idea. Let me let me go do that. And then

8:57like the other two are kind of iffy, but

8:59depends on the state, depends on the

9:01thing. And typically I do get something

9:02useful out of it. And then of course

9:04after I do that I again do a final pass

9:06before I submit it to humans to review.

9:08I really think it's super important that

9:09you review your own code. Um and do

9:11these passes before you submit it to

9:12another human. Otherwise it's just like

9:14super inefficient especially if the code

9:16is like mostly AI um written because I

9:19think this is a thing a lot of like

9:20juniors and people new to coding are

9:21doing especially vibe coding. They're

9:23like oh look I've built this thing. It's

9:24awesome. And they don't even bother to

9:26review their own code and it's just like

9:27a bad experience all around. So for more

9:29on that, you can check out this related

9:31post to review your AI's code. For my

9:33reviews, I'm typically just using cloud

9:34code or whatever's in the command line

9:36there. If it needs context across

9:38multiple files and then if it's like

9:40very isolated and like it's easy to just

9:42like copy paste into a chat, then I'll

9:44just use the chat cuz that's also pretty

9:45easy. But the main point is that I'm not

9:47using anything fancy. Um I have gotten

9:49my eye on things like Code Rabbit, which

9:51kind of like lives in the review tool

9:53and like reviews your code itself that

9:55way. Um and Claude's GitHub integration

9:58at work. We've got um codecs I think

10:00like connected to the repo and it's able

10:02to like do a review on on PR which I

10:04think is really cool and I probably will

10:06use those more in the future. I think

10:08they're really great for work and like

10:09organizations. I don't know if I'd use

10:11them in my side projects so much because

10:12it is just so fast to be like give me a

10:14review in the editor and get instant

10:16feedback that way. So not too bullish on

10:17this, but I also think you get a lot of

10:19benefits from just using pretty um

10:21simple workflows here. Next, so that's

10:22how I'm using AI these days for coding.

10:24Uh, similar processes also apply to my

10:27writing, although I do all the actual

10:28writing myself as I kind of find the AI

10:30writing to be flat and kind of soulless.

10:32My wife actually just wrote a blog post

10:34and um, I was reading it and I was like,

10:37are these written with AI? Like I feel

10:39like if you read enough you can kind of

10:40tell something is written with AI. And I

10:43also feel like a lot of the benefits

10:45that I get for writing these posts which

10:47is just like you know really solidifying

10:49what I'm thinking having things to look

10:51back on as like a snapshot of what I was

10:53thinking at that time and actually like

10:54sharing original thought so we can

10:56develop ideas together. A lot of those

10:58get lost if you let the AI do all the

11:00that thinking for you. And so that's

11:01something like I want to avoid at least

11:03in my own blog posts. But I still use it

11:05for research and reviewing blog posts. I

11:07think it's great for that. um giving

11:08ideas for like how to simplify sections.

11:11Like AI is excellent for that. I think

11:13more people should use it. Um and

11:15they're writing for that stuff. I would

11:16estimate that AI is probably writing

11:18around 40% of the code I submit.

11:20Although that code leans heavily on

11:22boiler plate. So things like unit tests,

11:24um things like plugging in the boiler

11:26plate to like some framework that you're

11:27using, stuff like that, but not not

11:29usually the the actual logic. And even

11:32that stuff is heavily edited or iterated

11:34on before pushing the code. you know, AI

11:36loves to submit all these like useless

11:38comments. Um, it is pretty verbose. It's

11:41not really good at like looking through

11:42a bunch of things and like simplifying

11:44them um to their core. And so like raw

11:4740% but also heavily edited 40%. So you

11:49can even say that it it was all AI,

11:51probably not. But I still think it's a

11:53significant chunk of code. Um, and it's

11:55kind of cool because it can kind of get

11:57the first draft out and then I think

11:58humans are pretty good at being like,

12:00um, oh, I see what you're doing there,

12:02but like it should probably be something

12:04else. And so you can kind of fine-tune

12:06it to be more closer to what you want.

12:08And I think that's like a pretty good

12:10combination of the two uh skills. Now,

12:12as for how much faster this makes me, I

12:14would probably say about 20% faster. You

12:16know, AI helps me research and code a

12:18lot faster, but it doesn't necessarily

12:19help things get over the line that much

12:21quicker just because you have to go

12:22through those extra layers of review um

12:25with the AI. But still, I think it's a

12:27pretty significant speed up. And I think

12:29that the outputs are generally just

12:32faster for the same or potentially

12:34slightly higher quality. Um, just

12:36because you are spending more time

12:38reviewing, you can try out more things

12:40with the AI because you're not like, uh,

12:41that's going to take an hour to code

12:43out. I don't want to do that. You just

12:44send the AI, it'll get it done in like

12:4610 minutes and then you can um, see if

12:47that made sense or go back. And so I

12:49think you're able to prototype in a lot

12:50more directions. Um, but still like

12:52overall throughput I'd say is probably

12:54around 20% faster. So if you liked your

12:57post then you might also like stop vibe

12:59coding start power coding how to write

13:01quality software faster with agentic AI.

13:03You might also be interested in I vibe

13:04coded a C# library with cloud code and

13:06here's six things I learned and finally

13:08how to checkpoint code projects with AI

13:10agents save your work keep projects on

13:12track and reduce rework. So that's it

13:13for this video. Thanks for watching and

13:15I'll see you in the next