YouTube2Text

Agentic Engineering in Grok Bot — Transcript

by Kun Chen · 4,457 words · 662 segments · language en · Watch on YouTube

Full transcript

  1. 0:00Hi, captains. I'm back with another
  2. 0:02video. This time showing you how I use
  3. 0:04my Grokbot setup to get work done. I'm
  4. 0:07just going to dive into my real setup
  5. 0:09here. So things may be a little bit
  6. 0:11messy, but I still think it's more
  7. 0:13useful to show you the real stuff
  8. 0:15instead of demos. Let's get into it.
  9. 0:17This is my Grokbot. I mainly use Grokbot
  10. 0:20for maintaining my open source projects.
  11. 0:23Um so if you haven't come across my open
  12. 0:25source projects yet, uh I have a bunch
  13. 0:28of open source projects. They are mostly
  14. 0:30agentic tools
  15. 0:31um and a lot of them got quite some
  16. 0:33popularity. So uh if you go into it,
  17. 0:35there's a lot of issues and PRs coming
  18. 0:37in.
  19. 0:38Uh maintaining these projects is not an
  20. 0:40easy effort. And Grokbot helped me so
  21. 0:42much with this. So we can see here I
  22. 0:46actually just asked Grokbot how much it
  23. 0:48helped me over the last month. It helped
  24. 0:51me process over 2,000 items
  25. 0:53split between issues and PRs. Um that is
  26. 0:56a lot of work. The only reason I could
  27. 0:59have kept those projects moving was
  28. 1:01because of my Grokbot setup here. So I'm
  29. 1:03going to walk you through how that
  30. 1:05works. Um I can probably show you a real
  31. 1:07example of how the automation works.
  32. 1:10Uh let me go to a smaller repo so it's
  33. 1:12less scary.
  34. 1:14Um we can see here in this repo there
  35. 1:16were issues and there were PRs, right?
  36. 1:19Uh if we go into a random PR here, um we
  37. 1:22can see this PR was raised by a
  38. 1:24contributor last week. If we scroll to
  39. 1:26the bottom here, we can see there is a
  40. 1:29comment from me
  41. 1:30speaking as Kuhn's first mate. Thanks.
  42. 1:33This is tidy change. And then there were
  43. 1:35a bunch of
  44. 1:37comments made, right? Like feedback on
  45. 1:39what the PR is supposed to do before
  46. 1:42it's safe to merge. Everything here came
  47. 1:45from my Grokbot, not myself directly.
  48. 1:47And pretty much every issue and PR in my
  49. 1:49open source repos gets a comment like
  50. 1:52this. And these comments guide the
  51. 1:54contributors through the process of
  52. 1:56getting their changes merged and getting
  53. 1:57their ideas implemented. If the PR is
  54. 2:00ready, it will give comments like this
  55. 2:02and uh tell the contributor what changes
  56. 2:04have to be made. And if the change is
  57. 2:06good and aligns with all the rules that
  58. 2:08I set for the repo, then FirstMate will
  59. 2:10merge the PR automatically. If the PR
  60. 2:12needs a decision from me, for example,
  61. 2:14it's a ambiguous product decision uh
  62. 2:16that's changing how things work, uh then
  63. 2:19it will escalate to me in Grok Bot. And
  64. 2:21I'll go back to Grok Bot now to show you
  65. 2:23how that works. Um so in Grok Bot, you
  66. 2:25can see on the left hand we have all
  67. 2:27these bots, right? Um I have one bot per
  68. 2:30open source repo. Uh so if I go to
  69. 2:32Treehouse, this is a bot that is
  70. 2:35specifically for triaging issues and PRs
  71. 2:38in my Treehouse repo. Basically, what
  72. 2:40this bot does is that uh it has a bunch
  73. 2:42of routines. There's a daily triage,
  74. 2:45there's a release process, and the
  75. 2:46cadence, uh how frequently this routine
  76. 2:49runs actually depends on how popular the
  77. 2:51repo is. So Treehouse is not a very high
  78. 2:54traffic repo, so a daily triage is
  79. 2:56enough. Uh for example, my FirstMate
  80. 2:59repo is a bit more busy, uh so it
  81. 3:02triages every 4 hours. So every bot is
  82. 3:05different here. Um and this triage will
  83. 3:08basically look at the open issues and
  84. 3:10PRs that haven't been triaged or have
  85. 3:12been updated since, and it will look
  86. 3:14into what's the right next step and make
  87. 3:16the comments I showed earlier uh or take
  88. 3:19actions to merge the PRs or close the
  89. 3:21issues. This is one example of an
  90. 3:23individual bot for one repo. I actually
  91. 3:25never come into this chat uh of this
  92. 3:28individual uh bot. I only talk to them
  93. 3:31through my FirstMate. FirstMate is the
  94. 3:33orchestrator agent that manages all
  95. 3:35these other agents for me, and it does a
  96. 3:37lot more than that actually. Uh I will
  97. 3:38show you in a bit, but I only talk to
  98. 3:40FirstMate. Uh you can see uh it's the
  99. 3:42only agent I talk to. I pretty much just
  100. 3:45close this uh all the time. And
  101. 3:47FirstMate uh will bring issues to me,
  102. 3:50bring things that needs my attention.
  103. 3:51For example, uh here it brought me uh
  104. 3:54captain FirstMate CI is hitting this
  105. 3:56limit. Okay, that's not good. Uh so that
  106. 3:59means uh the CI pipeline is probably
  107. 4:01failing on the first mate repo right
  108. 4:03now.
  109. 4:04Splits the slow jobs. Yeah. So it's
  110. 4:07suggested some potential solutions. I
  111. 4:10actually want to come back to this cuz I
  112. 4:11think this has repeated a few times. I
  113. 4:13want to solve this for good, not just
  114. 4:16keep patching this. But for now, I think
  115. 4:18you can get the idea that every
  116. 4:20individual bot's work is communicated
  117. 4:23through the first mate to me, and I only
  118. 4:25deal with first mate here. So problems
  119. 4:27from all these repos will surface in
  120. 4:29this single chat. I can just easily deal
  121. 4:32with them here. Many people actually
  122. 4:33already have a chief of staff kind of
  123. 4:35bot, right? But first mate is actually a
  124. 4:38little bit different. The first mate
  125. 4:39Grok bot started from the foundation of
  126. 4:41my open source project first mate here,
  127. 4:44which has been battle tested by a lot of
  128. 4:45users for many, many months now. And I
  129. 4:48think it's probably fair to say it's one
  130. 4:50of the projects that pushed the furthest
  131. 4:52on the chief of staff paradigm. It's got
  132. 4:54thousands of forks and a lot of active
  133. 4:56contributors. The first mate bot
  134. 4:58borrowed a lot of key learnings from
  135. 5:00this project on how to be an effective
  136. 5:02orchestrator. Here, let me show you some
  137. 5:04examples.
  138. 5:05I can go into the instructions of the
  139. 5:08first mate bot here. And if we go into
  140. 5:11context, memory, skills. Yeah, a lot of
  141. 5:14the instructions are in these skills.
  142. 5:16And one thing I want to show you is I
  143. 5:19think it's in the project management
  144. 5:21part. The first mate bot here actually
  145. 5:23uses a local SQL Lite database to manage
  146. 5:26the tasks. The benefit of this database
  147. 5:28is that first it's strongly persisted.
  148. 5:31So if the agents get reset or runs
  149. 5:33compaction, it doesn't lose anything.
  150. 5:35It's kind of like giving the bots a
  151. 5:37linear or Jira to work on complex
  152. 5:40projects.
  153. 5:42And the other thing that may be
  154. 5:43interesting to look at is in first mate
  155. 5:46charter.
  156. 5:47So there's a lot of instructions on how
  157. 5:49first mate manages the crew mates.
  158. 5:52One thing that may be interesting to
  159. 5:53point out is that when first mate
  160. 5:56notices that the crew mates are making
  161. 5:58mistakes
  162. 5:59or working inefficiently, it will update
  163. 6:02the behavior of those crew mates so that
  164. 6:04the crew don' make the same mistakes
  165. 6:06next time. This has been super
  166. 6:08beneficial uh and I can show you one
  167. 6:10example uh that I collected a while ago.
  168. 6:12This was from the First Mate bot
  169. 6:14summarizing what kind of improvements
  170. 6:16have been made to the other bots in the
  171. 6:18last two months and there's a lot of
  172. 6:20very concrete improvements on
  173. 6:22inefficiencies and noise. These
  174. 6:24improvements were made as the crewmates
  175. 6:26do their work and First Mate as the
  176. 6:28orchestrator and supervisor will just
  177. 6:31observe and make these tweaks so that
  178. 6:33the crew works more and more efficiently
  179. 6:35over time. First Mate also has uh a
  180. 6:38bunch of built-in uh engineering
  181. 6:40workflows uh like adversarial review uh
  182. 6:43lavish for planning. So basically when
  183. 6:45you import this uh First Mate uh bot
  184. 6:48into your Grok bot you get a whole stack
  185. 6:50not just a prompt. Maybe I can show you
  186. 6:52how I use First Mate to create this uh
  187. 6:54individual crewmates so I can just say
  188. 6:57create another triage crewmate uh for my
  189. 7:00new open source repo called headless.
  190. 7:05Yeah, that's it. If you're curious uh
  191. 7:08what I was using for voice input I use
  192. 7:10this tool called Handy. Uh I'm not
  193. 7:12affiliated with them or anything. Uh
  194. 7:14this is a free open source tool. I like
  195. 7:16to use an independent voice
  196. 7:18transcription tool instead of those
  197. 7:20voice inputs inside each app because
  198. 7:22then I can set a uh consistent key bind.
  199. 7:25I can have a consistent dictionary as
  200. 7:27well. Uh if you look at my Handy uh I
  201. 7:30think
  202. 7:31uh
  203. 7:32going to settings here.
  204. 7:33Uh yeah, it's really cool. You can just
  205. 7:35choose your own model like whichever
  206. 7:37model that works best for you. Um and I
  207. 7:40have this uh
  208. 7:43where is it? Advanced.
  209. 7:45Um yeah, I have this custom words set.
  210. 7:48These words are like really easy to mess
  211. 7:50up. Uh so when I add them here I can use
  212. 7:52them consistently everywhere uh which is
  213. 7:54great. Yeah, and now First Mate is
  214. 7:56creating another bot uh just like these
  215. 7:59ones. It said founded captain this is
  216. 8:01the repo. It has 20 open items uh from
  217. 8:04outside contributors. Uh yep, this is a
  218. 8:07new repo um which is really quite new.
  219. 8:09It's a It's an agent skill for setting
  220. 8:12up headless machines.
  221. 8:14And it's already got over 100 stars. I
  222. 8:16just published this a few days ago,
  223. 8:19which is awesome. So there's a lot of
  224. 8:20issues that I should probably take a
  225. 8:22look at. So once First Mate set up the
  226. 8:24triage crewmates, the triage crewmates
  227. 8:25can help me a lot with processing those
  228. 8:28issues. And here we can see the headless
  229. 8:30crewmate is already here, right?
  230. 8:32It's already starting to do things and
  231. 8:35it's being told by First Mate on what to
  232. 8:37do. Yeah, so this is going to be the bot
  233. 8:39that processes this headless
  234. 8:41repo for me moving forward. And it's got
  235. 8:44a routine now.
  236. 8:45It's doing the first pass triage as
  237. 8:47well. It's covering one PR and a bunch
  238. 8:50of issues. Yeah, so we'll just I'll head
  239. 8:53back to First Mate. I don't need to look
  240. 8:54at the crewmate there.
  241. 8:56But basically it's doing the triage. Now
  242. 8:59it's telling me that this repo has no
  243. 9:01vision.md.
  244. 9:02This is a very important concept I want
  245. 9:04to maybe touch on.
  246. 9:06So the reason I can let the bots
  247. 9:10process a lot of these PRs for me and
  248. 9:13know what is safe to merge.
  249. 9:16For example here, why did this bot make
  250. 9:19these comments and say these things are
  251. 9:21needed, right? These are not random. I'm
  252. 9:23not letting the LLM decide those things.
  253. 9:27Every repo I have has this vision.md
  254. 9:29file.
  255. 9:30And this vision.md file basically wrote
  256. 9:33down the principles for where I want to
  257. 9:36take this project, right? What are some
  258. 9:38of the boundaries? What are some of the
  259. 9:40principles? What kind of changes do I
  260. 9:41want to take? What kind of changes do I
  261. 9:43not want to take for the project? And
  262. 9:45every triage crewmate will respect this
  263. 9:48vision.md file and use this to determine
  264. 9:52what kind of PRs are safe to merge
  265. 9:54versus not. And that is in addition to
  266. 9:56good engineering practices. Yeah, and
  267. 9:58now this is telling me this headless
  268. 10:01repo does not have that yet. I actually
  269. 10:03should go write that down. So let me do
  270. 10:04that separately. And if you're curious
  271. 10:06how I write those vision.md files, this
  272. 10:08is the skill that I use. Um, this skill
  273. 10:11will actually push you really hard uh
  274. 10:14with hypothetical changes to the product
  275. 10:16and then test your boundaries to figure
  276. 10:18out what are the things that you are
  277. 10:20actually good with versus not. Uh and
  278. 10:22then after a few rounds uh it will help
  279. 10:25you write down your vision. I'll
  280. 10:26probably not wrap it whole into this
  281. 10:28right now. I might make another video at
  282. 10:30some point to introduce some of these
  283. 10:32important skills that I use. Uh if you
  284. 10:34think that's interesting, then let me
  285. 10:35know in the comments and I can
  286. 10:37prioritize based on what is interesting
  287. 10:39to people. Okay, back to Grokbot. Um we
  288. 10:42just started this headless crewmate,
  289. 10:44right, to triage the headless repo. And
  290. 10:47uh I think it's already done some work.
  291. 10:49We can see here two messages from
  292. 10:50headless. Uh this is um the headless
  293. 10:53triage bot sending messages to the
  294. 10:55firstmate bot and the firstmate relaying
  295. 10:57them to me. Uh this is useful because uh
  296. 11:00sometimes the bots will send something
  297. 11:02that doesn't require my attention and
  298. 11:04firstmate will just send something back,
  299. 11:06right, without bothering me at all. So
  300. 11:08although it may feel like oh the bots
  301. 11:10are sending messages to each other, uh
  302. 11:12is it wasting tokens? It's actually
  303. 11:14pretty useful. Um so now this is
  304. 11:16something that actually needs my
  305. 11:17attention. That's why firstmate uh send
  306. 11:19this message to me. Uh it says five of
  307. 11:22the open issues are real bugs. Okay. Uh
  308. 11:25the worst one is this uh on Windows. Uh
  309. 11:28there is an SSL setting. Uh okay.
  310. 11:30Uh this is another risk. Uh okay. This
  311. 11:34one PR number 25.
  312. 11:37Uh
  313. 11:38okay. This is a four line fix. That's a
  314. 11:40very easy fix to bring the code back in
  315. 11:42line with what the doc already promised.
  316. 11:44Okay. Uh so that's a correction of uh
  317. 11:47what we already want.
  318. 11:48Um
  319. 11:50um there is no grep tile. Okay, okay.
  320. 11:52Some of my other repos have grep tile
  321. 11:54enabled uh because grep tile um gives
  322. 11:56open source repos some free credits to
  323. 11:59go do code reviews, which I find pretty
  324. 12:01useful. Uh but on this repo I haven't
  325. 12:02set it up yet. Uh that's okay. But
  326. 12:04because I don't have the vision.md file
  327. 12:07written down yet, uh I want to be a
  328. 12:09little bit careful. So I'm going to
  329. 12:11actually look at this PR.
  330. 12:12Uh so this PR, yeah it's really just
  331. 12:15four lines.
  332. 12:16Um and the PR was raised through no
  333. 12:19mistakes, which we can see here. It's
  334. 12:22done a bunch of live validation to make
  335. 12:24sure it actually works. I pretty much
  336. 12:26all my open source repos require using
  337. 12:29no mistakes to raise the PR because then
  338. 12:31it ensures there's a minimum quality
  339. 12:33bar. So I can see this PR has already
  340. 12:35gone through a lot of rigorous internal
  341. 12:37steps before it's even raised here. So
  342. 12:41here we can see yeah a comment was
  343. 12:43already made by me which was coming from
  344. 12:46the triage bot.
  345. 12:48And it's saying CI is green, change
  346. 12:50looks good on the merits.
  347. 12:52Relaxes this. This is a captain decision
  348. 12:55hold. Okay. So because there's no
  349. 12:57vision.md, the bot actually think this
  350. 13:00behavior change requires my attention.
  351. 13:02So it said it's not waiting on the
  352. 13:04author.
  353. 13:05It's a captain decision hold. Captain
  354. 13:07means me. So that's why the message got
  355. 13:10to me and I'm here to decide whether
  356. 13:13this is the right thing. And I think I'm
  357. 13:15going to actually look at this.
  358. 13:17It's really just like correcting how the
  359. 13:20string matches
  360. 13:22and some documentation. So this is all
  361. 13:24good. I'm going to go back to the bot
  362. 13:26and say merge 25 and allow that going
  363. 13:28forward.
  364. 13:31So there we just triaged one PR, right?
  365. 13:33And first mate will continue on giving
  366. 13:36me other things that need my decision.
  367. 13:38So let's see what it comes up with. Also
  368. 13:40here you can actually see it's updating
  369. 13:42agent profile. So first mate is now
  370. 13:45updating headless to say it's okay to
  371. 13:47not have greptile now, right? So this is
  372. 13:51one of the things that happens when you
  373. 13:52have a first mate as the orchestrator.
  374. 13:55It just smooths out all these rough
  375. 13:57edges and make the crew mates work more
  376. 13:59and more smoothly down the road. Yep and
  377. 14:01first mate said the crew will merge it
  378. 14:03down and then immediately there's a
  379. 14:05message from headless saying it's
  380. 14:06already merged. Okay. So this basically
  381. 14:09this was done.
  382. 14:10If we go back to the PR, yeah it's
  383. 14:12merged. And
  384. 14:14I think we actually yeah a thank you
  385. 14:17note as well.
  386. 14:19So that's basically the automation
  387. 14:21handling this and occasionally the
  388. 14:23things the decisions will come to me and
  389. 14:26it's me reviewing some of the ambiguous
  390. 14:28stuff. Okay, now I'm curious what other
  391. 14:30decisions I have to make so I'm going to
  392. 14:32talk to first mate.
  393. 14:34What are some of the other PRs waiting
  394. 14:36for my decision?
  395. 14:38This is actually where first mate is
  396. 14:40going to read the database.
  397. 14:42If you see first mate is not talking to
  398. 14:44any crew mates right now.
  399. 14:46It's just reading the database and it
  400. 14:48will know which PRs are waiting for my
  401. 14:50decision
  402. 14:51without having to do a lot of messaging
  403. 14:53back and forth between the crew mates.
  404. 14:56Oh okay, this is actually a lot.
  405. 14:5888 PRs
  406. 15:00waiting for my word.
  407. 15:03Okay.
  408. 15:05Time sensitive. Okay.
  409. 15:08Okay.
  410. 15:10I'm going to yep, I'm going to say ahoy.
  411. 15:13Ahoy is a shortcut for giving me the
  412. 15:16most important thing
  413. 15:18one thing at a time. So let's see first
  414. 15:21mate I think it will actually do this.
  415. 15:24One thing waiting on you whether to go
  416. 15:26ahead with this one
  417. 15:28runs every PR checks on Ubuntu as well.
  418. 15:33Okay, it runs the PR CI pipeline on an
  419. 15:35additional OS version to just check
  420. 15:39whether it has problems. It's a
  421. 15:41contributor PR.
  422. 15:43I think this is good.
  423. 15:46But I'm curious why don't we block the
  424. 15:48PR when we know there's going to be a
  425. 15:50problem, right?
  426. 15:51Shouldn't we make this PR check actually
  427. 15:54block the PRs because it's going to be
  428. 15:56very soon that we will be forced to
  429. 15:58migrate.
  430. 16:01Okay.
  431. 16:02So yeah, sometimes these decisions come
  432. 16:04to me and I have to discuss with first
  433. 16:07mate a little bit.
  434. 16:09And first mate is checking how this PR
  435. 16:11does on that today to tell me
  436. 16:14based on source of truth. So yeah, let's
  437. 16:16see.
  438. 16:17Some of these decisions do need require
  439. 16:19some discussions.
  440. 16:20Um, so it says it's actually pretty
  441. 16:22quick. Um, captain, you are right. It's
  442. 16:24about to become required
  443. 16:26um since this now passes on this.
  444. 16:29Um, should we move to this? Yep, I
  445. 16:32actually think we should just move to
  446. 16:33this because it's going to be happening
  447. 16:36very soon. So,
  448. 16:37uh let's migrate. Okay. So, that's
  449. 16:39another decision made. Um, and I think
  450. 16:42after that I can just ask for another
  451. 16:43decision. Yeah, this is basically how I
  452. 16:45deal with the queue of all the issues
  453. 16:47and PRs that are coming to my open
  454. 16:50source repos. Yeah, it says the first
  455. 16:51mate crew has the rework of this
  456. 16:53underway. This is my rule that says
  457. 16:56basically if I have a slightly different
  458. 16:58idea for how a PR should be done, I will
  459. 17:02actually pay the cost of doing the
  460. 17:03modifications instead of asking the
  461. 17:05contributor to do that. Um, if it's a
  462. 17:08very big rework, then I'll send that
  463. 17:10back send the feedback back to the
  464. 17:12contributor.
  465. 17:13If it's just different opinions, I will
  466. 17:15do that. So, the crew will do the work.
  467. 17:18And if we look into the first mate bot
  468. 17:21here,
  469. 17:22it's viewing the PR. It's going to do
  470. 17:24something something. And then to make
  471. 17:26the code changes, it's going to use the
  472. 17:28cursor cloud agent. Yeah, you can see
  473. 17:30here launching cloud agent, right? So,
  474. 17:32it's using the cursor cloud agent to
  475. 17:34actually do the coding work. It doesn't
  476. 17:36do the work in this bot directly. Cursor
  477. 17:38cloud agent is a really good way to
  478. 17:41scale up a lot of this heavy coding work
  479. 17:44to a large amount of sandboxes. So, even
  480. 17:46if I have to make 100 PRs right now,
  481. 17:48this computer the Grokbot computer will
  482. 17:51not get too busy. So, I'm not going to
  483. 17:54like just sit here and wait. I'm going
  484. 17:56to say ahoy again. Basically just keep
  485. 17:59asking first mate to give me one more
  486. 18:01decision one more decision. And there's
  487. 18:04like 88 waiting for me. So, I'm not
  488. 18:06going to finish everything tonight, I
  489. 18:08don't think. So, yeah, I'm just here to
  490. 18:10show you how I work with first mate in
  491. 18:13Grokbot to do this open source
  492. 18:15maintenance, right? And you can see it's
  493. 18:17really beneficial to have one single
  494. 18:19first mate that can work across
  495. 18:21different projects for me. Otherwise, I
  496. 18:23will have to jump between all these
  497. 18:24bots, which is pretty much an
  498. 18:26anti-pattern uh in my eyes now. Oh, one
  499. 18:29thing I do want to go back to, which was
  500. 18:31that first mate uh CI was red, right? Uh
  501. 18:34we saw that earlier. Where is it? Yep,
  502. 18:36here. Um so I don't know why Grok bot
  503. 18:39will gray out some of these uh previous
  504. 18:41decisions. I think it's a new behavior,
  505. 18:43but basically I can just uh reply to
  506. 18:45this message and say uh what should we
  507. 18:48do about this now? Um actually I I want
  508. 18:51to show you something. Um
  509. 18:54ask the Kun bot to uh figure out how
  510. 18:57would Kun go about this.
  511. 19:01Uh
  512. 19:02there's a little uh typo here. Okay.
  513. 19:04Uh Kun bot uh is a really cool bot. Uh I
  514. 19:08will show you here. Um there's this bot
  515. 19:10called Kun. This bot basically has all
  516. 19:13my knowledge, uh everything I wrote in
  517. 19:15the public. Uh so my tweets, my uh blog
  518. 19:18posts, my YouTube videos, everything uh
  519. 19:21distilled into this bot. And the way it
  520. 19:24works is that uh I have this uh Kun
  521. 19:28repo.
  522. 19:29Uh you can actually find this as well.
  523. 19:31Uh it's a uh skill that you can use in
  524. 19:34any agent, not just Grok bot. It's a
  525. 19:36skill that is basically a near real-time
  526. 19:38distillation of my experience and uh
  527. 19:41knowledge and tools, workflows, and
  528. 19:42skills. Um the way this works is that in
  529. 19:45this uh repo, you can see this
  530. 19:48opinions.md.
  531. 19:49Uh these are actually my opinions. These
  532. 19:52are distilled from everything I said in
  533. 19:54the public, right? So um you can read
  534. 19:57this file. Uh it's really interesting. I
  535. 19:59I when I read it, I was very surprised
  536. 20:01how accurate it is. Um I I have a whole
  537. 20:04workflow um to uh generate this file
  538. 20:06every day based on what I said. So
  539. 20:09there's opinions, there's tools, uh what
  540. 20:12kind of tools that I use. Um and then
  541. 20:14there is uh assets content. Yeah,
  542. 20:17there's content as well. So every single
  543. 20:19video uh that I made, you can see the
  544. 20:22transcript uh written here. And
  545. 20:25basically, this is like everything I
  546. 20:27wrote or said or published in the
  547. 20:29public. It's all in this repo. And then
  548. 20:32the skill knows how to use all this
  549. 20:34content to answer questions. So
  550. 20:37basically, the more I write in the
  551. 20:38public, the more videos I make, the more
  552. 20:41blog posts I write, the more tweets I
  553. 20:43tweet,
  554. 20:44the more powerful this skill and the bot
  555. 20:46will be. So if I go back here, this bot
  556. 20:49is basically just using that skill.
  557. 20:51Let's see what First Mate said to me.
  558. 20:53The Coden Bot's answer is to leave the
  559. 20:5530 minutes limit alone.
  560. 20:58It's planned for fixing this at three
  561. 20:59parts, okay? Split the tests across
  562. 21:02parallel jobs.
  563. 21:04Yep.
  564. 21:06Warn when any job has its 20 minutes.
  565. 21:09Yep, that's something that I very often
  566. 21:11do. I set a lower threshold to get some
  567. 21:15warnings
  568. 21:16and so that we can actually fix things
  569. 21:18before they reach the hard limit.
  570. 21:21Speed up lint. Cool.
  571. 21:24These are actually like probably things
  572. 21:27I would do.
  573. 21:28And I likely already set those in some
  574. 21:31of my tweets in the public.
  575. 21:33Should we ship the Coden plan?
  576. 21:37Yes,
  577. 21:38I think that's a good idea. Basically,
  578. 21:40it's kind of interesting you see me
  579. 21:43talking to my bot,
  580. 21:44but this is actually pretty useful
  581. 21:46because I don't want to think about
  582. 21:48these things all the time, right? I
  583. 21:50already know how to solve some of these
  584. 21:52problems, so I just let the Coden Bot
  585. 21:54solve that. And you can use that too.
  586. 21:57You can just import the Coden Bot into
  587. 22:00your Grok Bot and ask it to solve like
  588. 22:03various kind of problems for you,
  589. 22:05especially when you think those problems
  590. 22:07are things that I may have some
  591. 22:09expertise on, then the bot can help you
  592. 22:11much faster than I could. Something
  593. 22:13that's even more fun is this Distill
  594. 22:16Anyone bot. This is another bot that I
  595. 22:18share that you can import into your Grok
  596. 22:21Bot as well.
  597. 22:22This bot can help you distill anyone a
  598. 22:26bot.
  599. 22:27So here I was playing. I distilled Mark
  600. 22:30Zuckerberg and Elon Musk and
  601. 22:33Sam Altman I think. And I asked them to
  602. 22:36discuss various kind of AI topics. It
  603. 22:39was really fun to watch. But basically
  604. 22:40this can be used to distill anyone that
  605. 22:43has public content. As long as they
  606. 22:45publish content publicly, then you can
  607. 22:48distill them into a bot and you can talk
  608. 22:49to them, ask them questions. This is
  609. 22:51very useful for getting a proxy to some
  610. 22:54of the domain experts that you would
  611. 22:56love to get inputs from. So you can very
  612. 22:58often just ask, hey how would that
  613. 23:00person solve this problem, right? Maybe
  614. 23:02just for the fun of it, let's distill
  615. 23:04someone else.
  616. 23:06Let's distill Jensen Huang. Okay.
  617. 23:11Let's see what happens. Okay, the
  618. 23:14distillation is happening.
  619. 23:16I hope it found the right person. I'll
  620. 23:18lock his identity from Wikipedia and
  621. 23:21check with you before saving. Okay,
  622. 23:23good. Okay, it's trying to confirm with
  623. 23:25me. Okay, this is
  624. 23:27founded Nvidia. Yep, that's him.
  625. 23:30So yeah. So now it's got the right
  626. 23:33person. It's now trying to actually
  627. 23:35fetch the content.
  628. 23:37Jensen Huang is locked. I'm setting up
  629. 23:39his folder, starting the first pass now.
  630. 23:41Cool. All right, it's done. Cool, taken
  631. 23:43from 22 talks and interviews and 21 of
  632. 23:47the posts. Okay,
  633. 23:48let's see. Okay, let's just talk to
  634. 23:51Jensen.
  635. 23:53What's on your mind? No, I think I will
  636. 23:55want to ask, how do you think about
  637. 23:59Chinese models?
  638. 24:02Let's see.
  639. 24:04Okay, it's got back an answer.
  640. 24:06Not the answer I was looking for though,
  641. 24:09but it's okay.
  642. 24:11One interesting thing is that this
  643. 24:14distillation process actually creates a
  644. 24:17voice profile.
  645. 24:18So this bot will actually talk like
  646. 24:21Jensen.
  647. 24:22And that's something that's really cool.
  648. 24:25You can see here,
  649. 24:26look, you have to start with the facts.
  650. 24:29Something like This is actually similar
  651. 24:31to how Jensen talks, right? So, this is
  652. 24:34something that's really cool, and I hope
  653. 24:35you all will have fun with this. All
  654. 24:37right, that's it for today. I will share
  655. 24:40links to all these Grok bots that I
  656. 24:41mentioned, so you can import them into
  657. 24:43your Grok bot as well. All of them are
  658. 24:45free. I'm not trying to sell anything
  659. 24:47here, just trying to share something
  660. 24:49that I have been using and hope you will
  661. 24:51find them useful as well. Thank you for
  662. 24:53watching and see you next time.

About this transcript

This page contains the full transcript of Agentic Engineering in Grok Bot by Kun Chen, generated from the public captions YouTube serves with the video. The transcript has 4,457 words across 662 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.