YouTube2Text

My AI Coding Workflow That Actually Works — Transcript

by Chris Titus Tech · 4,455 words · 603 segments · language en · Watch on YouTube

Full transcript

  1. 0:00This is my AI development workflow. It's
  2. 0:03not perfect and it's constantly
  3. 0:05evolving. And this is why I think you
  4. 0:07you don't see very many people talk
  5. 0:09about this, but I do encourage people to
  6. 0:11make comments on this video uh and just
  7. 0:14say, "Hey, this is how I do it
  8. 0:15differently compared to yours. This is
  9. 0:18meant for bigger software, more complex
  10. 0:22software. Obviously, you need to
  11. 0:24understand Git and those types of basics
  12. 0:27before you get into a workflow like this
  13. 0:29one, but it will give you really good
  14. 0:33code. Uh, and I'm going to walk through
  15. 0:35it. Now, if you want to just be like,
  16. 0:37hey, I want just one prompt and have it
  17. 0:39loop a whole bunch and then just spit
  18. 0:41out something,
  19. 0:42watch my last video about how I spend
  20. 0:441.5 billion tokens and realize that's
  21. 0:47not really a good idea, especially on a
  22. 0:49big project. you have to have manual
  23. 0:51gates. Uh so having said that, let's get
  24. 0:54into my workflow right here. Uh a broad
  25. 0:57overview for this is you're going to
  26. 0:59have an agents, a spec, a road map. Now
  27. 1:02the agents can be machinewide. Spec and
  28. 1:05road map are repository specific. So you
  29. 1:08typically make a git repository and
  30. 1:10you'll put those spec and roadmap files
  31. 1:13in. And then you plan out and build
  32. 1:16these files and also kind of understand
  33. 1:19as uh before you start really coding or
  34. 1:22having LLM code, you're going to have
  35. 1:24and instruct it specifically to make uh
  36. 1:28a testing suite for you and then check
  37. 1:30that testing suite. And then once you
  38. 1:32have like the scaffolding or the the
  39. 1:34shell, the guard rails, if you will in,
  40. 1:37then you can actually build. A lot of
  41. 1:40people skip this entire thing or just go
  42. 1:43I'm gonna go in plan mode and then hit
  43. 1:46implement plan. And I think that's how a
  44. 1:48lot of people use LLMs and that's why
  45. 1:50they get poor results. So how what I did
  46. 1:54is I built my own repository called
  47. 1:56Titus-ai. It has an agents, a spec, a
  48. 1:59road map all with examples in them. And
  49. 2:02then from there um I pulled specific
  50. 2:06things depending on some of the leaders
  51. 2:08in in the space. I think Cararpathy's
  52. 2:09agents uh is public and I took that and
  53. 2:12I made slight modifications to my tastes
  54. 2:15of how I like to code or or develop
  55. 2:18projects and then have just sample
  56. 2:20examples and then also a skills folder.
  57. 2:23But the skills again, don't get too
  58. 2:26caught up on these things. And remember,
  59. 2:27they're all plain text. So when you look
  60. 2:30at like um the agents file, look at
  61. 2:33these and if there's a rule in here you
  62. 2:35don't agree with, delete it. Add your
  63. 2:37own rules. This is pretty plain text and
  64. 2:40it should follow this. Um, but like skip
  65. 2:43flattery, filler, ceremon openings,
  66. 2:46emojis, whatever it is. Put this in here
  67. 2:48and say do not use this. I want straight
  68. 2:51working code only. I want you to be as
  69. 2:54specific and minimal and only touch the
  70. 2:57things that the task requires. I don't
  71. 2:59want you to just start going off on a
  72. 3:01tangent. It's important to get these
  73. 3:04principles. And also I do have some
  74. 3:08token reduction. I know I need to get
  75. 3:10more into that space with token
  76. 3:12reduction. I've tried other projects
  77. 3:13like Headspace and uh many others and uh
  78. 3:16token reduction can be a good thing, but
  79. 3:19often it does lead to a little bit
  80. 3:21poorer results. So depending on how much
  81. 3:23token reduction you're doing, uh you
  82. 3:26might reduce your output, but it can
  83. 3:29solve. But don't get stuck on that side
  84. 3:31of things. I I'm only briefly touching
  85. 3:33on here. If you're interested in the
  86. 3:34token specific thing, I typically do
  87. 3:37targeted token reduction things like
  88. 3:40giant log files and those types of
  89. 3:42things. I'll use RTK to cut down on. But
  90. 3:45I won't use RTK on every task that the
  91. 3:47LLM does. Uh because often I find it
  92. 3:50just repeats the task and increasing
  93. 3:52tokens used, which is the cost of this
  94. 3:54thing. But I could go through this
  95. 3:56entire file, but we're going to be here
  96. 3:58all day for that. I just wanted to kind
  97. 3:59of just show these are the rules. The
  98. 4:01spec is in the actual repository itself
  99. 4:04when so if we go into I think my website
  100. 4:07has a spec. Uh it'll actually kind of
  101. 4:10specify hey we're going to be using Hugo
  102. 4:121.62
  103. 4:14uh it specifies the version. A lot of
  104. 4:16people when they build out their
  105. 4:18projects using an LLM LLM will pick old
  106. 4:21crap. They will pick old versions things
  107. 4:23with security problems all that. It's
  108. 4:26just not good. So specify pin the
  109. 4:29version that you're on right now using a
  110. 4:31spec file. And then I also go in I think
  111. 4:34I even have a linting uh thing kind of
  112. 4:36just showing the structure of this. I
  113. 4:39have some automation scripts that fire
  114. 4:41off through GitHub actions as well on
  115. 4:42this website. And I thought I'd put I
  116. 4:46probably should add something here. So I
  117. 4:50I probably also validate using markdown
  118. 4:52lint. Uh there's actually something a
  119. 4:54package on this specific system called
  120. 4:56markdown lint CLI 2 and it would verify
  121. 4:59all the markdown so there wouldn't be
  122. 5:01any formatting errors as well. So that
  123. 5:03would be an idea. So like when you're
  124. 5:05going through and doing the validation
  125. 5:06you're like oh okay I need to do that.
  126. 5:08Uh another issue like with uh my other
  127. 5:11project where I have my entire uh guey
  128. 5:15that I'm building. I decided you know
  129. 5:17what screw it I'm going to build my
  130. 5:18whole desktop environment for Linux
  131. 5:20video coming out sometime soon. um using
  132. 5:24QuickShell. QuickShell had specific
  133. 5:26tasks and differences with uh its
  134. 5:29linting. So when it was linting the QML
  135. 5:32files using QT6's base one, it wasn't
  136. 5:35being able to grab the types and things
  137. 5:37like that. So it was not classifying it
  138. 5:39correctly. So then the linting would
  139. 5:41throw errors which would blow up the
  140. 5:43code. It just not good. So I went ahead
  141. 5:45and made something in the spec file to
  142. 5:47say, "Hey, make sure you're always doing
  143. 5:49this with the types and then run your
  144. 5:51lint and then it will actually show
  145. 5:53correctly." So anytime you have a
  146. 5:54specific spec or something that needs to
  147. 5:57change, think of it that way. Build the
  148. 5:59test suite, understand how to validate,
  149. 6:02say, "This is where uh you're going."
  150. 6:04And I kind of write this down. I think
  151. 6:06it's right in here. It shows kind of
  152. 6:08like the the purpose of the project and
  153. 6:10all the things in the agents. But in the
  154. 6:11spec here, you need the problem,
  155. 6:14intended users, uh required behavior,
  156. 6:18user experience, architecture, major
  157. 6:20components is probably a bigger one. I
  158. 6:22probably should clean this up a little
  159. 6:23bit. Security and privacy requirements,
  160. 6:26which it kind of goes to the versioning
  161. 6:28I was talking about. And then the big
  162. 6:30one here is acceptance criteria. I
  163. 6:33always in that acceptance criteria
  164. 6:35always have it go through some kind of
  165. 6:36test or validation. You got to let it
  166. 6:38know when it succeeds. And then for road
  167. 6:42mapping, this is where you would
  168. 6:44actually say here's how you'd probably
  169. 6:46go about doing this. Now, uh you just
  170. 6:49need to make sure you give it that hey
  171. 6:51at phase one it will be complete when
  172. 6:54this happens kind of uh workflow. And
  173. 6:57then you you break that up and chunk it
  174. 7:00into different phases. And in each
  175. 7:01phase, I like to do a giant review. But
  176. 7:04to showcase where I'm going with this
  177. 7:09on DWM, here is an example run. I
  178. 7:13already have my testing. I already have
  179. 7:15all the specs, road maps, every kind of
  180. 7:17everything logged out. But I want to
  181. 7:19change
  182. 7:21this to show the audio percent next to
  183. 7:24the audio icon. add
  184. 7:27uh audio percent next to audio icon. And
  185. 7:34what this will do is it'll go through
  186. 7:38find the skill. So it found my quick
  187. 7:40shell skill. I also put where there's
  188. 7:44spec and skill. There's some overlap
  189. 7:45there. So whenever it hits quick shell
  190. 7:48skill, it will actually go ahead and
  191. 7:50grab that markdown CLI I just went over.
  192. 7:52So, just know that the skills can be
  193. 7:54modified as well to where whenever you
  194. 7:57hit bash and your bash is doing
  195. 7:59something weird, you might change that
  196. 8:01up a little bit. You can do it in the
  197. 8:02spec sheet for the project or the skill.
  198. 8:05So, now I can go into any project with
  199. 8:07quick shell and it'll just remember my
  200. 8:08tastes uh based on that skill that I
  201. 8:11made. Now, I think there's some
  202. 8:15terminology changes here too between
  203. 8:17claude and codeex. Claude, I think,
  204. 8:19refers to them as workflows where Codeex
  205. 8:22kind of refers to them as skills. And I
  206. 8:25I've noticed a lot of um
  207. 8:30guides and things I've seen online, it's
  208. 8:33still such a fresh thing. Don't take
  209. 8:35this as gospel. What I'm saying today
  210. 8:37will probably be wrong tomorrow or
  211. 8:40change, but it's just understanding,
  212. 8:42hey, this is kind of what each phase
  213. 8:45should look like. So the the phrasing
  214. 8:47I'm using might change, but the purpose
  215. 8:51of what I'm doing does not. So you
  216. 8:53always have these tasks that you're
  217. 8:56doing. And it's just so important uh to
  218. 8:59do them because if you're just trying to
  219. 9:01do one task and just say run wild, just
  220. 9:04don't keep going until it's done. It's
  221. 9:07just going to use tons of money, tons of
  222. 9:09tokens, and then probably not even
  223. 9:10produce anything good. add to live
  224. 9:12install and make a PR with these
  225. 9:16changes.
  226. 9:17So, this is going to add it to my live
  227. 9:19install, restart my quick shell and then
  228. 9:24create a pull request that can then be
  229. 9:27verified. So uh even though this is done
  230. 9:31or well it should be done once the
  231. 9:32restart happens here the poll request
  232. 9:35will open up and then we start to
  233. 9:37approach the next phase that would
  234. 9:39happen with any kind of implementation
  235. 9:41and that next phase is the review phase
  236. 9:43which is really important. This is where
  237. 9:45unit testing happens. This is also where
  238. 9:48um I typically pull independent
  239. 9:50reviewers. Uh I've used code rabbit.
  240. 9:52I've used uh just straight up at codeex
  241. 9:56review. You can do claude review. Um,
  242. 9:59but it needs to be outside of this
  243. 10:01agent. And that agent is so important to
  244. 10:05do. And you can see right here, it
  245. 10:07already reloaded, rated why I'm why I'm
  246. 10:10live here. So, within the last minute,
  247. 10:11it's went ahead and changed that. But
  248. 10:13we're also making that pull request
  249. 10:14because we need to verify the manual
  250. 10:17gate. Well, we can see right here. Yeah,
  251. 10:20it did work. But how did it implement it
  252. 10:22is what I'm now wondering. like, okay,
  253. 10:25that was magic. That was pretty cool.
  254. 10:27That would have taken me hours before,
  255. 10:29which took, you know, maybe 2 minutes.
  256. 10:32But now we need to verify. And this is
  257. 10:34where the last video I went into how it
  258. 10:37fell apart on me. This is where it all
  259. 10:40happens. The magic happens at that
  260. 10:41manual gate, which we're about to hit.
  261. 10:44As this live starts, it's like, okay,
  262. 10:46everything looks good. It now displays
  263. 10:48the 100%. Fantastic. Now, I want to see
  264. 10:51the poll request. And you'll also notice
  265. 10:54it's taking screenshots of the screen
  266. 10:56here and then verifying that it shows
  267. 10:59there. So you can actually live verify
  268. 11:01automatically.
  269. 11:03But I also caution you have that manual
  270. 11:05gate the poll request at each phase with
  271. 11:08that criteria we wrote at the road map
  272. 11:12is what I'm looking for. I don't trust
  273. 11:15AI enough to just manually rubber stamp
  274. 11:18this stuff. So important to have that
  275. 11:20manual gate. So here we go. And while
  276. 11:22that PR is getting made, there's another
  277. 11:25part of the setup we need to talk about,
  278. 11:27too. I talked briefly about that
  279. 11:30versioning. LLMs always pick old
  280. 11:32versions that have tons of security
  281. 11:34problems in them. They've gotten better
  282. 11:35about not leaking AI API keys and stuff,
  283. 11:38but it's still not perfect. So, I always
  284. 11:41enable dependabot and code QL. Oh, I
  285. 11:45heard it. It looks like that PR is now
  286. 11:47ready. The cool thing about this is it
  287. 11:52will also be checking to see if there's
  288. 11:54any updates because LLMs love to choose
  289. 11:56old stuff and we want to make sure we're
  290. 11:59using like an LTS version. If it's using
  291. 12:00node, it needs to be node 24. You know,
  292. 12:04that's the current LTS as the making of
  293. 12:06this video. Now, you might not know that
  294. 12:08off the top of your head, so you're
  295. 12:09going to need to look some of this up,
  296. 12:11but it's really important to have the
  297. 12:14security of Depinabot and QXL in there.
  298. 12:18So, I usually go code scanning and then
  299. 12:20check the projects. Enable code scanning
  300. 12:22on all your project, I would say, just
  301. 12:25to see these security errors. It's so
  302. 12:27important to do and this is free, so why
  303. 12:29wouldn't you do it? So, here we go. We
  304. 12:32got that. Currently, I have Code Rabbit
  305. 12:34hooked up. um not a sponsor yet. Maybe
  306. 12:38maybe in the next video. So, what I'm
  307. 12:40going to do is I'm going to just take it
  308. 12:41from not ready to ready for review. And
  309. 12:44I'm not going to merge this yet because
  310. 12:46I need independent verification. So, I'm
  311. 12:49going to trigger the review here. And
  312. 12:52then also we could do an at@codeex
  313. 12:56review. So, let's go at codeex review
  314. 13:00comment. And this is where AI is cool to
  315. 13:03me. I don't really care. so much about
  316. 13:05it creating the code. That's neat to put
  317. 13:07100% on the volume button. I could doing
  318. 13:10that myself. But where it's really neat
  319. 13:13is code review. This is where I use a
  320. 13:16lot of my data. Like when I'm looking at
  321. 13:18what I'm using, uh we just had a reset
  322. 13:21today. I'm already 25% of my max plan.
  323. 13:24So at a $100 plan, uh doing this type of
  324. 13:29stuff, it's going to be heavy. If you're
  325. 13:30a $20 a month user on Codeex, it's gonna
  326. 13:35eat through that pretty fast. So, uh,
  327. 13:38that's another thing to consider about
  328. 13:40doing heavy development work. You're not
  329. 13:43going to be able to do it without the
  330. 13:44$100, $200 subscriptions. And I think
  331. 13:47that's another part of this that needs
  332. 13:49to be talked about. I think uh, we will
  333. 13:52be transitioning to local models for
  334. 13:54some, like I know the new Quinn models
  335. 13:56coming out that can run on a lot of
  336. 13:58local machines. It's not going to be as
  337. 13:59good as these models, but I think it
  338. 14:02using this type of workflow with that is
  339. 14:04feasible sometime in the future. Um, and
  340. 14:08I'm know that this type of thing at
  341. 14:11$100, $200 is also going to be highly
  342. 14:14subsidized. I know I'm using way more
  343. 14:16compute than the cost that I pay them.
  344. 14:19So, I'm getting a hell of a deal right
  345. 14:21now, but I know this deal is not going
  346. 14:22to last forever. So, kind of think about
  347. 14:25that as well because even if you take
  348. 14:28the LLM out of this, this is just a good
  349. 14:31workflow for developing software
  350. 14:33regardless of using an LLM or not. I
  351. 14:36know I've said that a couple times, but
  352. 14:37I just can't emphasize that enough. So,
  353. 14:39review. Now, if you're doing this with a
  354. 14:42human, you'd have multiple humans
  355. 14:43looking at this project. Uh, but as you
  356. 14:46see, a lot of my usage is actually
  357. 14:51all code review. I use a ton of code
  358. 14:54review. I love code review. And you can
  359. 14:57manually trigger this with just an at
  360. 14:59symbol. You'll see the eyes. That means
  361. 15:01it's hooked up properly. If you don't
  362. 15:02see codeex do the emote after doing a
  363. 15:05review, that means it did not uh did not
  364. 15:09recognize it. So yeah, that means you
  365. 15:14need to go back into your your settings,
  366. 15:16whether it's GitHub or another one, and
  367. 15:18then hook it up properly. Also, another
  368. 15:21thing I like to do as well is local code
  369. 15:24rabbit. They have a CLI baked in to
  370. 15:27where you can actually before you make
  371. 15:28the poll request, if it's a bigger
  372. 15:30change, I'll like to do a couple local
  373. 15:32passes with the review before sending it
  374. 15:34up and making a pull request. Again,
  375. 15:37very re review heavy, very testing
  376. 15:39heavy. And a lot of the code written in
  377. 15:42this, if we look at the actual code, oh
  378. 15:45my god, this seems pretty excessive, but
  379. 15:48look at the code that it's made.
  380. 15:5116 lines added, five lines removed. But
  381. 15:55if you look at this, this is 11 lines of
  382. 15:58code. We're using a lot of compute for
  383. 16:00those 11 lines of code. But what it
  384. 16:03yields
  385. 16:04is so good that I'm like, hey, this is
  386. 16:08fantastic. It's only taken me a couple
  387. 16:10of minutes. We're using my monthly plan
  388. 16:13and it's going to town. This is the
  389. 16:17future of development in my eyes and how
  390. 16:19I work on my thing. But let's let's see
  391. 16:22what the code review comes back with.
  392. 16:24Usually, it will find something on the
  393. 16:27code review. Uh, this one already found
  394. 16:29a couple things. You can see uh hide the
  395. 16:32percent when audio is unavailable. Yeah,
  396. 16:35that's probably a good idea to do. uh
  397. 16:38record the new panel behavior in the
  398. 16:40change log. This adds a user visible
  399. 16:44panel, but the diff updates only the QML
  400. 16:48and regression assertion leaving
  401. 16:50unreleased section and change log
  402. 16:52without an entry for the new volume. So
  403. 16:54this is a documentation issue. Uh both
  404. 16:57are valid though. So let's let's just
  405. 17:00have AI do those things for us. So what
  406. 17:03I'll say is address and resolve PR
  407. 17:07review feedback and commit and push.
  408. 17:11Commit adds a commit locally. So another
  409. 17:15thing about understanding git is it'll
  410. 17:17add it locally and say here's a
  411. 17:19checkpoint essentially for your code and
  412. 17:21then a push will push it to wherever
  413. 17:23that code repository resides. In this
  414. 17:25case it's github.com. And then it's
  415. 17:27going to go through that thread, come
  416. 17:30back, and sometimes the reviews I I read
  417. 17:33it and I'm like, "That doesn't make
  418. 17:34sense to me." Or, "Ah, that's a little
  419. 17:36bit too verbose. I don't really care
  420. 17:38about the change log. No one's going to
  421. 17:40look at this." All those things. You can
  422. 17:42actually make the call here. This is the
  423. 17:44manual gate I'm talking about. You're
  424. 17:47looking at the code. You're looking at
  425. 17:49the comments from the reviews from AI.
  426. 17:52And this is the assistance. It's meant
  427. 17:53to be a pair program or just like having
  428. 17:55an expert sitting right here next to me
  429. 17:57and going, "Hey, Titus. Yeah, uh, you
  430. 18:00idiot. You You forgot about what happens
  431. 18:02if there's no audio. It's just going to
  432. 18:03have a weird percent up there with just
  433. 18:06random symbols. What what what are we
  434. 18:07doing?" And it's those things that
  435. 18:11that's that's the power of it. It's not
  436. 18:14that it wrote the percent. It's the fact
  437. 18:16that it reiterated, reviewed it a whole
  438. 18:19bunch, and then thought of all these
  439. 18:21other cases that I I didn't think up or
  440. 18:25I didn't realize until I would put it
  441. 18:26out in the wild and then a bunch of
  442. 18:28people would review it or use it and
  443. 18:30then go, "Oh, I don't have audio." And
  444. 18:32then all of a sudden there's a bunch of
  445. 18:33weird symbols up there and then they
  446. 18:35would make an issue on GitHub and then I
  447. 18:37would see that issue and then I go, "Oh,
  448. 18:39yeah, I forgot to add that." And then I
  449. 18:41would manually code that in. So that was
  450. 18:43the old method and it's essentially just
  451. 18:46like the new method. It's just a a it
  452. 18:49looks a little different and it's way
  453. 18:51faster and it reiterates a lot quicker.
  454. 18:54So that's where the changes happen.
  455. 18:56That's where the magic of this is. I
  456. 18:58just I I just want to beat that into
  457. 19:00people's heads because often I I hear a
  458. 19:03lot of AI haters just say, "Hey, it why
  459. 19:06don't people talk about hallucinations
  460. 19:07anymore? Why don't two people talk about
  461. 19:09like it's more about the user and less
  462. 19:13about the model. It's less about what's
  463. 19:16going on. It's it's how you use it. And
  464. 19:19and that's the big thing here. You you
  465. 19:22got to know these basic things to test
  466. 19:26it properly to check and fix and then
  467. 19:29understand how the development should
  468. 19:31look. And you can see our commit in real
  469. 19:33time come through. Uh, and there it is.
  470. 19:38Now it'll actually do a review on that.
  471. 19:41You see it's going through my continuous
  472. 19:43integration here and going through all
  473. 19:45the tests. It's doing smoke tests on
  474. 19:48each distribution, Arch, Debian, and Red
  475. 19:51Hat Enterprise Linux. You're going
  476. 19:53through all this. Oh my god, I can't
  477. 19:55tell you how cool that is. Instead of me
  478. 19:57having to install and do all these
  479. 19:59different tests or manually write these
  480. 20:01tests, it would come through and and
  481. 20:04should check out all these code raex.
  482. 20:11You could probably go ahead and take a
  483. 20:13look at it, but I want to see exactly
  484. 20:15the changes made to address that
  485. 20:17feedback right here. If v unavailable,
  486. 20:21go ahead and make this non-visible.
  487. 20:23Okay, that makes sense. Then we have the
  488. 20:26test that was changed to verify that.
  489. 20:29And then in the change log, add the
  490. 20:31volume percent to address all of our
  491. 20:34reviews. And I even got the little ping
  492. 20:37from Herder. This is like my first day
  493. 20:39on Herder. So, uh, probably a video
  494. 20:41coming out soon for that. But we're
  495. 20:43good. Everything looks great here. I am
  496. 20:46happy with those code reviews. We have
  497. 20:48the approval. Wait for my last smoke
  498. 20:51test. This is actually standing up a
  499. 20:52whole Docker. probably a little
  500. 20:54excessive for this type of uh setup, but
  501. 20:56this is just GitHub CI that I'm burning
  502. 20:59through. I think we can safely say this
  503. 21:01is going to go through. I don't we
  504. 21:02didn't really have any
  505. 21:05distribution or distro specific changes
  506. 21:07here. So, you you could technically want
  507. 21:10to wait for all your CI. Now, I will say
  508. 21:14doing it this method, you're not going
  509. 21:16to move as fast. you're going to move
  510. 21:17way faster than you did a month ago or
  511. 21:20for me a month ago, but for you maybe a
  512. 21:22couple months ago when or couple years
  513. 21:24ago when LLMs just kind of sucked um at
  514. 21:29writing code, but now you are to a spot
  515. 21:33where you can move so fast you just make
  516. 21:36a bunch of crap. But I encourage you to
  517. 21:40slow things down. Make a YouTube video
  518. 21:43about this process. kind of go through
  519. 21:46and you can kind of see, hey, what's
  520. 21:48happening as it goes through all your
  521. 21:51testing? Does that make sense to you?
  522. 21:53Uh, are your reviews coming back? Are
  523. 21:55you addressing and resolving those
  524. 21:57reviews? And then this this is great. We
  525. 22:01now have a volume percent. I actually
  526. 22:04wanted to add this. I just saw it and I
  527. 22:05was like, "Oh, that'd be a good test of
  528. 22:08this workflow." And then we're done. And
  529. 22:10then I'm just going to delete the
  530. 22:12branch, prune it off, and we are on
  531. 22:14main. Uh, now DWM Titus has a percent in
  532. 22:18the QML. We looked at the code, we
  533. 22:20looked at the reviews, we addressed all
  534. 22:21the changes, we've documented all the
  535. 22:23changes. It makes sure everything's nice
  536. 22:26and pretty, and it's way better than it
  537. 22:28was beforehand because a lot of times I
  538. 22:30just would like, ah, documentation, I'll
  539. 22:31do that later. And now it forces me to
  540. 22:34do it now. And it's really good. So that
  541. 22:37was my final merge gate. And I wanted to
  542. 22:40kind of give you a real world example.
  543. 22:41This I put an article down below uh my
  544. 22:44AI workflow. This is due to change. I'll
  545. 22:46probably update this article as I go.
  546. 22:48But remember, slow is smooth, smooth is
  547. 22:50fast. Don't move too fast.
  548. 22:55Wait for that testing to get polished
  549. 22:59and come back green.
  550. 23:02Wait for those reviews to happen. You
  551. 23:04might have to wait 5 10 minutes for
  552. 23:06Codex to review. I've even had Codex
  553. 23:08take upwards of 20 or 30 minutes to
  554. 23:10review a big PR. So, think about
  555. 23:14chunking those into smaller poll
  556. 23:16requests so it happens a little bit
  557. 23:17quicker so the work for you're not
  558. 23:19trying to do too much too soon. Uh and
  559. 23:22and follow this guide. Even if you're
  560. 23:25not using an LLM, this is kind of how
  561. 23:28you should be doing things regardless.
  562. 23:31Have good development workflow. This
  563. 23:33will make good software for you. and
  564. 23:35it's made great software for me. Better
  565. 23:38software than I made before LLMs really
  566. 23:41became good. And they're good enough
  567. 23:43now. Don't be waiting for the next
  568. 23:45model. Don't be waiting for mythos
  569. 23:48fabled ultra plus $1,000 a month plan.
  570. 23:52That's not going to fix it for you.
  571. 23:55This is here today and it's going to
  572. 23:57work way better than that will in a
  573. 24:01year. It's not going to fix you. It's
  574. 24:04not going to fix your bad on development
  575. 24:08workflows. It can't just magically read
  576. 24:11your mind, even though sometimes it does
  577. 24:14feel like that happens. And with that,
  578. 24:16let me know your comments. What did I
  579. 24:18miss? What do you do that I didn't touch
  580. 24:20here? Uh, I know I didn't miss. Uh,
  581. 24:22there's still plugins like Super Power
  582. 24:24and some other stuff I needed to kind of
  583. 24:26sprinkle in, but this is already getting
  584. 24:28really long in the tooth and I wanted to
  585. 24:30make a more succinct video about it. Let
  586. 24:33me know. I know this stuff. I see a
  587. 24:35bunch of AI. Oh, I hate AI or it's going
  588. 24:37to ruin the world and all the politics
  589. 24:39and this is where it's headed. I don't
  590. 24:41talk about any of this. I'm just saying
  591. 24:43this is what can practically be done and
  592. 24:46that's all I care. What can it do for me
  593. 24:48today? And as soon as it's not doing
  594. 24:51that job or it's too expensive, I kick
  595. 24:53it to the curb. I am a practical guy, I
  596. 24:57get done. And that's what I'm known
  597. 24:59for. So, I encourage you to do the same.
  598. 25:03Take what you will out of this. Take the
  599. 25:05good bits. Tell me the bad bits because
  600. 25:07I always like and I like refining this
  601. 25:10policy to make it as good as possible.
  602. 25:12And with that, I'll see you in the next
  603. 25:15one.

About this transcript

This page contains the full transcript of My AI Coding Workflow That Actually Works by Chris Titus Tech, generated from the public captions YouTube serves with the video. The transcript has 4,455 words across 603 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.