How I Keep Up with AI Without Drowning in It (Free Skill) — Transcript

Full transcript

1. 0:00If you've never felt overwhelmed with
2. 0:01all the news in the AI space, then you
3. 0:03should just go start your own YouTube
4. 0:05channel and teach us your ways. Cuz I
5. 0:07can guarantee that's not the case. At
6. 0:09least some of the time, maybe even right
7. 0:10now, you're feeling overwhelmed with the
8. 0:12onslaught of new models and product
9. 0:13releases, viral tweets, and GitHub
10. 0:15repositories that you have to check out.
11. 0:17I mean, every single week it feels like
12. 0:18there are multiple things that come out
13. 0:20that change everything. And so it
14. 0:22becomes overwhelming. And the big
15. 0:23question we have to ask ourselves is,
16. 0:25how do we keep up with it all? And how
17. 0:27do we figure out what's fluff versus
18. 0:29what is actually worth our time looking
19. 0:31into and maybe incorporating for
20. 0:33ourselves. And that's the big question
21. 0:35I've been answering for myself this
22. 0:36entire year, building what I call my AI
23. 0:39content engine within my second brain.
24. 0:41And so in this video, I'll show you how
25. 0:43this whole system works and even better,
26. 0:45how you can build it yourself pretty
27. 0:46easily using a skill that I have for
28. 0:48you. And I'm excited for this because
29. 0:50this system is a lifesaver for me. It
30. 0:52tracks hundreds of sources on my behalf.
31. 0:55And then for all of the announcements
32. 0:57and blog posts and articles, everything
33. 0:58that goes through the system, it filters
34. 1:01based on what it knows I care about in
35. 1:02my content. So all I get every single
36. 1:04day is a concise and easy-to-read daily
37. 1:07digest that covers everything I should
38. 1:10care about. So it's not helping me track
39. 1:12all the AI news, because even for a
40. 1:14digest, that would be a lot to read. But
41. 1:15it's just the things that I care about.
42. 1:18And I've made this system very accurate
43. 1:20over time. So this filter truly is
44. 1:22taking out the fluff for me and helping
45. 1:24me hone in on what's going to make my
46. 1:25content better, my AI coding workflows
47. 1:28better, just the things that are
48. 1:29personal. So within my main skills
49. 1:31GitHub repository, I have our skill for
50. 1:33today, build signal engine, signal
51. 1:35content. I call my system either. This
52. 1:37workflow guides your coding agent
53. 1:39through building the exact same system
54. 1:41that I have that's saving me hours and
55. 1:43hours a week of going through all the AI
56. 1:45news. And the reason why I have a skill
57. 1:48instead of just giving you the system as
58. 1:50an open-source repo you set up right
59. 1:52away, is I want you to be able to
60. 1:53customize things as you build it from
61. 1:55the ground up. And so as a part of this
62. 1:57process at the very beginning, it grills
63. 1:59you. It figures out what sources you
64. 2:01want to track, how you want to fetch
65. 2:02them, the filtering you want to do, like
66. 2:03what you actually care about, and then
67. 2:05it builds the system. But, the value in
68. 2:08this skill is that all of the best
69. 2:09practices I've learned, you know,
70. 2:11lessons I've learned the hard way,
71. 2:12refinements that I've made as I've been
72. 2:14using my content engine for over 6
73. 2:16months now, that is all built into this
74. 2:18skill. But, before we go through this
75. 2:20skill together, I want to show you how
76. 2:22the system works, because I'm a firm
77. 2:23believer for anything we set up or run
78. 2:25ourselves, we should truly understand
79. 2:27how it's working. So, the first thing we
80. 2:29have to figure out before we set up any
81. 2:31of the filtering or teaching of what we
82. 2:32care about, is we need to build the data
83. 2:35layer, the connections to the different
84. 2:36platforms, and figuring out what we're
85. 2:37going to pull into the content engine in
86. 2:39the first place. There are so many
87. 2:42sources that I'm pulling in every single
88. 2:44day. Reddit, YouTube, RSS feeds, Hacker
89. 2:46News, GitHub repositories, change logs
90. 2:48from my favorite platforms like Cloud
91. 2:50Code. There are hundreds of posts that
92. 2:52I'm feeding into my content engine every
93. 2:54day, but it's also able to work through
94. 2:56everything very, very quickly. So, a ton
95. 2:58of sources, but I made it very
96. 3:00efficient. And of course, that's part of
97. 3:01what the skill will help you build as
98. 3:03well. So, the skill is going to ask you
99. 3:04the different sources you want to
100. 3:05connect to, and then also how you want
101. 3:07to do so. Because for every single
102. 3:09platform, there are many different ways
103. 3:11we can connect, directly with an API or
104. 3:13a CLI tool or MCP server. And so, I've
105. 3:15done a lot of that for, you know, like
106. 3:17GitHub and archive and Hacker News. But,
107. 3:19something that I've started to use
108. 3:20recently is Appify. It is a single
109. 3:22platform that allows me to very easily
110. 3:24connect my agents to thousands of
111. 3:26different sources, like YouTube and
112. 3:28Reddit. Also, the kinds of things where
113. 3:30usually the APIs are more restricted.
114. 3:32And so, again, it's not like you have to
115. 3:34use the same solutions that I'm using,
116. 3:36but this is what's working really well
117. 3:37for my content engine, and the skill
118. 3:39will help you customize. So, you figure
119. 3:41out your data layer and what you want to
120. 3:42connect to, and then we move on to the
121. 3:45rest of the process where we're actually
122. 3:46setting up the filtering and getting to
123. 3:48that daily digest. So, first thing we do
124. 3:50when we run the content engine is we
125. 3:52pull all of our sources. And I run my
126. 3:54content engine every single day. But,
127. 3:56the very first thing we have to do
128. 3:57before we do any kind of processing or
129. 3:59filtering is deduplicate. So, storing
130. 4:01everything in a central database and
131. 4:03keeping track of the announcements and
132. 4:05blog posts and things that we have
133. 4:06already seen. Because we don't want to
134. 4:09have Jev and our large language model
135. 4:12process the same blog post or
136. 4:13announcement or whatever twice. Cuz
137. 4:15that's just a waste of our tokens. And
138. 4:17so, before we get into any of the model
139. 4:19usage, we want to do a deterministic
140. 4:21filtering and deduplicating as much as
141. 4:23we possibly can. It's part of what makes
142. 4:25this whole system very fast and token
143. 4:28efficient. And so, we deduplicate. And
144. 4:30then I also have some code filters, some
145. 4:32deterministic filters that I've built
146. 4:34in. And so, if there are really old
147. 4:36posts that are likely stale, I'm just
148. 4:38going to scrap that right away. Or if
149. 4:39there's a lack of engagement on a post.
150. 4:41Like if there's, you know, for example,
151. 4:43a Hacker News article or a YouTube video
152. 4:44I pulled in that only has a couple of
153. 4:46hundred views. I mean, maybe that's good
154. 4:48information, but a good general rule of
155. 4:50thumb that I've seen is usually I can
156. 4:52skip past that thing. And so, being able
157. 4:54to track that before I'm actually using
158. 4:56my tokens to determine if it's important
159. 4:58to me. And then for whatever gets past
160. 5:00the deterministic filters and
161. 5:01deduplication, I send that into the
162. 5:04first model call using Jev. So, this is
163. 5:07something I covered in my last Second
164. 5:08Brain Jev Use Cases video. I'll link to
165. 5:10that right here. But, I love using Jev
166. 5:13as the first line of defense for models
167. 5:15doing the filtering. And so, Jev hasn't
168. 5:18been around for super long. Before Jev,
169. 5:20I was using LLMs to filter everything.
170. 5:22But, that meant making an LM call for
171. 5:24every single source that made it past
172. 5:27the first couple of stages. And you can
173. 5:29use a cheaper model for this. Like
174. 5:30Haiku, especially with the new Haiku
175. 5:325.5, you can have a lot of reasoning
176. 5:35capability for super cheap. But, I still
177. 5:37don't want to use an LLM as much as I
178. 5:39possibly can avoid it. And so, that's
179. 5:41why I'm using Jev as the first line of
180. 5:43defense. Now, I did try a version of
181. 5:46this system where I didn't have an LLM
182. 5:48at all and Jev was the last line of
183. 5:51defense. My accuracy actually went down
184. 5:53a lot. I had a lot of things slip
185. 5:55through that I didn't care about, a lot
186. 5:56of things filtered out that I actually
187. 5:58did, but I found this to be the sweet
188. 6:00spot here. So, I use Jev to make more of
189. 6:03the obvious decisions. Anything that is
190. 6:04very obviously fluff or very obviously
191. 6:07something I care about, it's going to
192. 6:08make that decision first, right? So, if
193. 6:11it's like, "Yes, we definitely should
194. 6:12drop this." then it exits and I never
195. 6:14have to make an LLM call.
196. 6:15But for anything where it's unsure or it
197. 6:17knows that I need to definitely keep,
198. 6:19that's when we send it to an LLM to just
199. 6:21get everything ready for the digest and
200. 6:23do any further filtering that I need.
201. 6:25And so, to give you some concrete
202. 6:26numbers here, for this run that I did
203. 6:28with 864 posts, it only costed 3 cents
204. 6:32for all the calls I made to Jev and then
205. 6:35for what survived to Claude Haiku, it
206. 6:37only costed 42 cents. And so, it's less
207. 6:40than half a dollar every single day. Not
208. 6:43that bad for how much time it saves me,
209. 6:45it is definitely worth it. Now,
210. 6:46obviously, the price for you is going to
211. 6:48change depending on how many sources
212. 6:50you're pulling from and how much text
213. 6:52you get per source. And so, also
214. 6:54something you want to be very aware of
215. 6:55here is for every source you pull
216. 6:57through, generally, you don't want to
217. 6:58just dump the raw output of the MCP
218. 7:00server call or API call into Jev or your
219. 7:03LLM. There's almost always going to be
220. 7:05some things you can deterministically
221. 7:07strip like headers and footers, just
222. 7:09extra spacing and things that the agent
223. 7:11doesn't really need. That's one of the
224. 7:12reasons I like using Apify because for
225. 7:14all the runs that I have here, as it's
226. 7:16pulling from my different sources, it
227. 7:17automatically returns markdown that is
228. 7:19optimized for the agent, like just the
229. 7:21information that it needs. And so, using
230. 7:23something like Apify, you have to worry
231. 7:25about it less, but if you do have a
232. 7:27custom setup, just keep in mind it's
233. 7:28very important for you to establish up
234. 7:30front and of course, that's one of the
235. 7:32things that my scale helps you with. And
236. 7:34I also put together this little demo
237. 7:35here just to make things really
238. 7:36concrete. So, this is not my full
239. 7:38content engine. This is just from
240. 7:40pulling through a couple of subreddits,
241. 7:42but it's a nice visualization to see
242. 7:44what it looks like to filter some things
243. 7:46with Jev and then send other things to
244. 7:48the LLM. And so, for every single source
245. 7:51that makes its way to Jev, it's going to
246. 7:53evaluate based on three pieces of
247. 7:55criteria. This is built into the skill
248. 7:57as well. It's going to figure out, is
249. 7:58this on topic? Is this something Cole
250. 8:01would care about? Is this actually new
251. 8:03and concrete or is it just a toy demo? A
252. 8:06lot of things that I find with blog
253. 8:09posts and especially subreddits is
254. 8:10people are just sharing little things
255. 8:12that they built to try to benchmark
256. 8:14models that I don't really care about.
257. 8:15And so, this is one of those lessons I
258. 8:17learned where I see these like toy demos
259. 8:19that I like to call them a lot. And I
260. 8:20just want to filter that out right away.
261. 8:22So, anything that scores even reasonably
262. 8:24high with this like a 0.7, I'm just
263. 8:25going to filter it out. I'm just going
264. 8:26to drop it. I'm not going to send that
265. 8:28to the LLM. And so, there's some things
266. 8:30that are obvious enough where like this
267. 8:31is so good. Like the announcement for
268. 8:33Sonnet 5.5 or the release of the
269. 8:36official 1.0 of Pi. Like yes, I
270. 8:38definitely want to keep that. I don't
271. 8:40even need the LLM to decide. And then
272. 8:42other things where it's more nuanced and
273. 8:43maybe I want the LLM to dive a little
274. 8:45bit more into my memory system cuz Jev
275. 8:48can't really do that. That's when I send
276. 8:50it off to the LLM to reason more, do I
277. 8:52actually keep it? The system is still
278. 8:54light. It's still going to be cheap when
279. 8:56it goes to the LLM, but just not quite
280. 8:58as much. And then for every single one
281. 9:00of these sources that I want to bring
282. 9:02through the content engine, I just
283. 9:03manage them all with a single
284. 9:05configuration JSON file. It could not be
285. 9:07easier for me to manage and extend this.
286. 9:10Like here are all of the RSS feeds I
287. 9:12want to follow, all the new blog posts
288. 9:14from this, YouTube channels, GitHub
289. 9:16repositories, the categories for archive
290. 9:18research papers, for Appify, I have the
291. 9:21searches I want to perform every time
292. 9:22with Reddit and web search and YouTube.
293. 9:25So, catching more general things outside
294. 9:27of the individual channels as well. All
295. 9:29of this together makes it so I feel like
296. 9:31I have a really good comprehensive grasp
297. 9:33on the whole internet as new things are
298. 9:35coming out. Awesome. So, now that you
299. 9:37understand the system at a high level,
300. 9:38we can get into using the skill to build
301. 9:40your own personal content engine. I'll
302. 9:43just walk you through getting started
303. 9:44cuz the skill really does go through the
304. 9:45entire thing with you. And so, the
305. 9:47easiest way to install this skill is to
306. 9:50follow the instructions in the readme in
307. 9:52my skills repository. I'll link to this
308. 9:54in the description. If you're using
309. 9:56Claude Code, I have a Claude Code plugin
310. 9:58to make it so incredibly easy to bring
311. 10:00this into your system. And if you're
312. 10:02using a different coding agent, you can
313. 10:03just point it at the URL for this
314. 10:05repository and ask it to set up the
315. 10:07skills or just download the single skill
316. 10:09that we're using right now. And so, once
317. 10:11you have this set up, I would recommend
318. 10:13running the skill in a new repo, a new
319. 10:15fresh folder where you want to build the
320. 10:17content engine. You can optionally
321. 10:20specify the path to a separate
322. 10:22repository. Like if you're building it
323. 10:23in your second brain, but you want it to
324. 10:25actually be created somewhere else. I
325. 10:26just wanted to leave that as an option,
326. 10:28but generally when you build this,
327. 10:30you're simply going to do /build signal
328. 10:32engine and that is it. You don't even
329. 10:34have to specify any arguments. So, I'm
330. 10:36going to send this in and the first
331. 10:38thing it's going to do, like I mentioned
332. 10:39earlier, is grill you with a bunch of
333. 10:41questions so it can really be custom to
334. 10:43you. So, it starts by asking where
335. 10:45should we actually build this? It's
336. 10:46going to default to recommending where
337. 10:48we currently are. So, I'm just going to
338. 10:49go with that because I created a fresh
339. 10:50folder for this. And then it'll start
340. 10:52getting into asking, you know, what do
341. 10:54you actually need to keep up with,
342. 10:55right? So, the higher level questions so
343. 10:57it can even maybe recommend different
344. 10:58sources. So, I want to keep up with, you
345. 11:00know, AI tools that I use or maybe you
346. 11:02want to keep up with specific platforms
347. 11:04you want to call out. It's entirely up
348. 11:05to you how you answer these questions
349. 11:07and communicate with it. Tell me about
350. 11:09something in your AI tools that you
351. 11:10found out about late or missed entirely.
352. 11:12Right? Like what is the actual problem
353. 11:15if we get more granular here that we're
354. 11:16trying to solve. Right? So, maybe it's
355. 11:18like later we found out about something
356. 11:20in the change log that we wish we knew
357. 11:21about a week earlier or we found out
358. 11:23from a YouTube video. Like, "Oh man,
359. 11:25this thing came out a couple days ago
360. 11:26and I didn't find out until this YouTube
361. 11:28video. Like I really wish I knew about
362. 11:29it right away."
363. 11:30And then getting more into the sources
364. 11:31and everything, I'm not going to go
365. 11:33through and answer every single question
366. 11:35here cuz I want you to do that very
367. 11:36custom to what you care about, but I'll
368. 11:38show you the final output. So, I went
369. 11:40through this process already where I
370. 11:41answered every single one of the
371. 11:43questions. It really does grill you. You
372. 11:44have to be a little bit patient here,
373. 11:46but I I recommend really investing the
374. 11:48time into this up front because it's
375. 11:50going to make it so that I can just rip
376. 11:51through the entire implementation. So,
377. 11:53you spend, you know, 10-15 minutes
378. 11:55answering its questions and then it
379. 11:56builds out the entire thing and it
380. 11:58actually does it in phases so that it's
381. 12:00very comprehensive and even testing
382. 12:02things for your personal engine as it's
383. 12:05building it. So, going back to the
384. 12:06diagram here, after the interview, the
385. 12:08skill going to walk itself autonomously
386. 12:10through a six-stage process to build out
387. 12:12everything. So, we're not just trying to
388. 12:14one shot it. We're having it split up
389. 12:16into bite-size pieces of work and it's
390. 12:18going to test after every single one of
391. 12:20them. It would definitely fall flat on
392. 12:22its face if it just tried to one shot
393. 12:24everything. So, it starts very simple,
394. 12:25just one source that you agreed upon in
395. 12:27the interview, building that end-to-end
396. 12:29and testing it. Then, moving on to
397. 12:31building every other source and fully
398. 12:33establishing this data layer here and
399. 12:35doing it according to the platforms and
400. 12:37methods that you specified. Like for
401. 12:39myself for this demo here, I said that I
402. 12:41want to use Appify for everything and so
403. 12:43that's what it's going to be integrating
404. 12:44with as it builds out the data layer.
405. 12:47But, whatever you decided, whatever
406. 12:48customizations you have, it's going to
407. 12:49work with that. If you want to use APIs
408. 12:51directly, whatever, it's going to build
409. 12:53that out and then it's going to build
410. 12:54the code filters, the deterministic ones
411. 12:56and test that, the decision steps with
412. 12:58Jev and the LLM, building the format for
413. 13:00the digest, and then also helping you
414. 13:01set the schedule for the frequency you
415. 13:03want to run this on. And so, all that
416. 13:05together is just going to rip through
417. 13:07this. And so, if we go back to my
418. 13:08conversation here to the very end, boom,
419. 13:10we have the full engine built. Now, this
420. 13:12might not be perfect right away, so you
421. 13:14can definitely iterate with the agent
422. 13:15here, but it is going to be a very good
423. 13:18starting point. The other thing that I
424. 13:19built into the skill is it produces a
425. 13:21profile at the end. So, you have an
426. 13:23at-a-glance view for the sources you're
427. 13:24pulling from, the filters and decision
428. 13:26questions that you have. So, if there's
429. 13:28something that doesn't look perfect here
430. 13:29or exactly what you want, you can just
431. 13:31adjust it and then point the agent at it
432. 13:33and say that like, you know, here's a
433. 13:34new profile. I want you to implement the
434. 13:36differences. And so, this is kind of
435. 13:38your way to just constantly adjust
436. 13:39things. So, keep in mind, the content
437. 13:41engine is not reading this every time it
438. 13:43runs, but whenever you want to change
439. 13:45something, you can just update it here
440. 13:47in natural language and then tell the
441. 13:49agent to implement those changes. So,
442. 13:51there you go. That is how my content
443. 13:52engine works and how you can build it
444. 13:54yourself with the skill. And even if you
445. 13:56don't build it from scratch here and you
446. 13:58just want to take a couple of golden
447. 13:59nuggets from this video and incorporate
448. 14:01into your own process, I hope that you
449. 14:02find that really useful. And so, if you
450. 14:05appreciated this video, you're looking
451. 14:06forward to more things on agentic coding
452. 14:08and building up your second brain, I
453. 14:10would really appreciate a like and a
454. 14:12subscribe. And with that, I will see you
455. 14:14in the next video.