I Tested the Viral Reverse Engineer Anything Skill. What You Need to Know. — Transcript

Full transcript

1. 0:00So this GitHub repo right here, reverse
2. 0:01engineer anything, has been absolutely
3. 0:03blowing up. Matthew Bourbon himself said
4. 0:04software is done. Now he's kind of
5. 0:06joking a little bit. This narrative has
6. 0:07been around for, you know, over a year
7. 0:09since vibe coding has really taken off.
8. 0:11But there is some truth to it because
9. 0:12the idea of reverse engineer Anything is
10. 0:14that you see a feature you like on some
11. 0:16sort of app or software that already
12. 0:17exists and then you can have your agents
13. 0:19using the frameworks inside of here,
14. 0:21understand how that functionality works
15. 0:22and then actually just go ahead and
16. 0:24recreate it. So today I wanted to
17. 0:25quickly talk about how this actually
18. 0:26works and why I don't think this is
19. 0:28anything really that new and because of
20. 0:30that how you can use this and also the
21. 0:32right mindset to be able to actually
22. 0:34build things way better. So I've been
23. 0:35playing around with this REA repo and
24. 0:37I've built games, I've built software,
25. 0:39I've built different tools and apps and
26. 0:41just tested out a bunch of things and it
27. 0:43is very impressive. But as I was going
28. 0:44through and testing it out and using it,
29. 0:46I realized that it's not anything new
30. 0:48and we have probably been doing this for
31. 0:49a long time. I definitely have been
32. 0:51doing this for a long time. The way that
33. 0:52I think about it is that AI is not
34. 0:54really that creative. It can't really
35. 0:55think outside the box. But if you give
36. 0:57it a very clear vision, it can execute
37. 0:58on that for you really well. And if we
38. 1:00take a look at some examples that I've
39. 1:01actually built a while before this
40. 1:03GitHub repo existed, you'll see what I
41. 1:06mean. On the left, we have my human
42. 1:07edited Instagram reels. I'm going to
43. 1:08mute this real quick so I can actually
44. 1:10play it. But you can see the vibe,
45. 1:11right? You've got B-roll, you've got
46. 1:13different font colors, you've got
47. 1:14shimmers, you've got different types of
48. 1:15animations. And what I did in order to
49. 1:17make a skill that now edits my Instagram
50. 1:19reels, is I had it go watch like 30 of
51. 1:22my best performing Instagram reels. And
52. 1:24now we get the style on the right, which
53. 1:25looks pretty much identical. The same
54. 1:26sorts of different colors and fonts and
55. 1:28vibe and animations and all of this kind
56. 1:31of stuff. And now this just lives inside
57. 1:32of a skill. And the point is what I had
58. 1:35to do here was I couldn't say, "Hey, I
59. 1:36want you to build me an Instagram
60. 1:37editing, you know, skill that has this
61. 1:40sort of vibe and does these sort of
62. 1:41transitions." No, what I did is I said,
63. 1:43"Hey, look at these videos. analyze all
64. 1:45of these frame by frame and understand
65. 1:47what's going on under the hood. And
66. 1:48that's essentially me decompiling my
67. 1:50human edited real style. Here's another
68. 1:52example with a motion edit. On the left
69. 1:53was one from Jordan Watkins, and you can
70. 1:55see that it's really, really well done.
71. 1:57And he said, "I'd love to see AI try and
72. 1:58recreate this." So, I was like, "Huh,
73. 1:59let's see what it can do." So, I
74. 2:01downloaded the video, gave it to Opus,
75. 2:02and said, "Hey, analyze this frame by
76. 2:04frame, understand what's going on and
77. 2:06how it works and why it works like this,
78. 2:07and then build me a skill around it and
79. 2:09show me that you can edit a video
80. 2:10similar to it." The one on the left is
81. 2:11still better with motion design and
82. 2:13sound design. It still is a little bit
83. 2:14better, but the thing is Opus could not
84. 2:16have thought of this. This is such a
85. 2:17creative design that I don't think
86. 2:19Claude or any AI could have created it,
87. 2:22but because I'm showing it, this is what
88. 2:23I want. Analyze it now build it for me.
89. 2:26It was able to do a much better job. And
90. 2:27when I've talked to you guys about how
91. 2:29to build the best skills, this is
92. 2:30exactly what I preach. I preach reverse
93. 2:32engineering the output. If you want to
94. 2:33build a Friday report, have it look at
95. 2:35five examples of good Friday reports
96. 2:37because now it knows exactly what you
97. 2:39want. I've talked a lot about how
98. 2:40communicating your intent has to be
99. 2:42specific. If you give an agent a generic
100. 2:44goal, it's going to give you a generic
101. 2:46result. So, this whole idea of reverse
102. 2:48engineering anything and having it
103. 2:50specifically look at an app that you
104. 2:51want to recreate and then copying the
105. 2:53features from it, that is way more
106. 2:55specific. Essentially, what you've done
107. 2:57is you've given the agent a very, very
108. 2:58specific PRD, but instead of you having
109. 3:00to type out the exact PRD with all the
110. 3:03specs, you're just saying, "Look at the
111. 3:05app. That's your PRD." I've done this
112. 3:07with landing pages, you know, taken
113. 3:09elements or animations or designs from
114. 3:11other landing pages. Here was an example
115. 3:12of one that I did with Glido. I've done
116. 3:14this with apps, right? Like you can see
117. 3:16right here, I've got this thing um right
118. 3:18here, which is basically the Loom clone
119. 3:20that I made from using this repo, right?
120. 3:22Like I had it look at Loom. I had it
121. 3:24understand how it can work with like the
122. 3:25different bubbles. You can see it does
123. 3:27the recording countdown. It might not
124. 3:29work because my face cam is recording
125. 3:31right right somewhere else. So, I'll
126. 3:32play a clip real quick. All right. All
127. 3:33right. So, I just wanted to test out how
128. 3:35this actually works to make sure that
129. 3:38this is like a decent software. So, I'm
130. 3:40going to see if I can pause this. And
131. 3:42now we can go ahead and keep on rolling.
132. 3:44I can move this thing around. I can also
133. 3:46make it smaller or larger while I'm
134. 3:48recording. That's very nice. And let's
135. 3:49go ahead and stop and see what happens.
136. 3:51But the thing that's important to
137. 3:52realize is that we didn't clone Loom.
138. 3:54This is a very different looking app. We
139. 3:55didn't clone like all of the exact UI
140. 3:58and the design elements. We cloned the
141. 4:00functionality. We understood how does
142. 4:02the little bubble work? How can you move
143. 4:04it around? How can you change the window
144. 4:05that you're recording, right? Like we
145. 4:07talked about this sort of stuff and we
146. 4:08had it analyze that sort of stuff. Loom
147. 4:10obviously records and stores your things
148. 4:12in the cloud in Loom's cloud, but this
149. 4:14one is just local. So, it's storing
150. 4:15these files in my own folder inside of
151. 4:17this app called Herk Recorder. I also
152. 4:19pointed this at Sand Castles and had it
153. 4:21inspect the functionality and understand
154. 4:23and research how it works and what it
155. 4:25does. And then it gave me basically my
156. 4:27own version of Sand Castles and I just
157. 4:29had to connect my API key. So it helps
158. 4:30me analyze Instagram content and analyze
159. 4:33my own content. And you can also see
160. 4:34this is true with something like our AIS
161. 4:36eval studio which lets us actually eval
162. 4:38our automations. We took a lot of
163. 4:40inspiration here from Langfuse and we
164. 4:41understood how they work, how they show
165. 4:43things, how they run the things. And now
166. 4:44we have our own version right here. And
167. 4:46it's not a carbon copy of course, but we
168. 4:48had it look at and grab some of the best
169. 4:51core features and how they work and why.
170. 4:54Real quick guys, I've got this
171. 4:55completely free website design skill
172. 4:56that I'm giving away. I use this for
173. 4:58every single website that I build,
174. 4:59including our AI automation society
175. 5:00site, which I think is pretty slick. And
176. 5:03I absolutely love this website. It's
177. 5:05called ScrollCraft, and it understands
178. 5:06things like scroll driven animations and
179. 5:09layering, but also it has things in
180. 5:10there like taste, typography, spacing,
181. 5:13depth. It's just a really good website
182. 5:15design skill in general, and it's
183. 5:16completely free. So, the link for this
184. 5:17will be down in the description. But
185. 5:19let's get back to the video. So, I guess
186. 5:20the point I'm trying to make there is
187. 5:21reverse engineering has been around for
188. 5:23a long, long time. But the reason why
189. 5:25this repo is blowing up is because this
190. 5:26repo gives your agents a very practical
191. 5:29and thought through framework for
192. 5:31decompiling and iterating in order to
193. 5:33actually create what you want to create.
194. 5:35And if you haven't been using your
195. 5:37agents in that way where you show them
196. 5:38the end goal and then help them sort of
197. 5:40work backwards to get there to build the
198. 5:42skills to give you a better output, then
199. 5:44you definitely should start thinking
200. 5:45about this not just when you're trying
201. 5:46to build apps or software, but when
202. 5:48you're just driving your day-to-day and
203. 5:50you're building skills. So, let me just
204. 5:52break down real quick, real simply what
205. 5:54actually this is and how it works. So,
206. 5:56reaver reverse engineer anything. It
207. 5:58helps your AI agent study something and
208. 6:00then help you build it. So, let's take a
209. 6:02look at like a recipe, a bakery example.
210. 6:05So, let's say the bakery has a cake that
211. 6:07you love. You order it every week and
212. 6:09you love it and you wanted to reverse
213. 6:10engineer the cake to get the recipe
214. 6:12because the recipe is locked away. It's
215. 6:13the the bakery's IP and they don't give
216. 6:15it away. So you buy a cake, the bakery
217. 6:17ships the end result to you and they use
218. 6:20the recipe to make that obviously. Now
219. 6:22you inside of your kitchen, you can
220. 6:23actually, you know, tear open the cake,
221. 6:25you can taste it, you can try to look at
222. 6:27it under microscope and you're
223. 6:28essentially trying to reverse engineer
224. 6:30the recipe from that cake. And this is
225. 6:32the process of decompiling, which means
226. 6:33you're actually tasting it. You're
227. 6:35inspecting it. You're trying to break it
228. 6:36down to figure out what are, you know,
229. 6:39the different code blocks and what is
230. 6:40the actual letters in here that make the
231. 6:43cake. And there's where you get sort of
232. 6:44like your own sort of trial and error
233. 6:46where you're building and you're
234. 6:47comparing and you're trying to iterate
235. 6:48on it until you hit the end result. So
236. 6:52what happens is if we take that example
237. 6:53of Loom, you can ask your agent, hey,
238. 6:55use reverse engineer anything and go
239. 6:57figure out how Loom's camera bubble work
240. 6:59so we can work that into our own, you
241. 7:01know, screen recording app. Your agent
242. 7:02would then inspect the files and inspect
243. 7:04the functionality of the app. It would
244. 7:06look through the evidence. It would
245. 7:07build its own version and then you
246. 7:08iterate until it basically matches up in
247. 7:10terms of functionality. And like I said,
248. 7:12it comes out so well because you're
249. 7:13pointing it at an end result. You're not
250. 7:15saying, "Hey, make me a cake." You're
251. 7:16saying, "Make me this very specific
252. 7:18example of cake that tastes like this
253. 7:20with this much density and it has, you
254. 7:22know, this type of frosting and this
255. 7:23many layers." It's just way more
256. 7:25specific of a request. So, this whole
257. 7:26doc that kind of walks through and
258. 7:28explains some of this, I will link in my
259. 7:30free school community for completely
260. 7:31free if you guys want to grab it, as
261. 7:33well as this doc, which just kind of
262. 7:34goes through like how to get the most
263. 7:36out of it. Because what I found after
264. 7:37building a bunch of stuff is it's so
265. 7:39hard to just say, hey, go reverse
266. 7:40engineer Loom. It's easier to pick out
267. 7:43features. So you use Loom for a little
268. 7:44bit and maybe it's like, okay, I want to
269. 7:46build this myself. What are the best
270. 7:48features of Loom that I like? Or, you
271. 7:49know, I want to rebuild my CRM. What are
272. 7:51the features of, you know, Salesforce or
273. 7:53whatever I use that I really, really
274. 7:54like? Okay, now we use our agent to go
275. 7:56pick out those specific features,
276. 7:57investigate them, decompile them, and
277. 7:59then bring them over to our own app. And
278. 8:01then what I did kind of for fun is I had
279. 8:03cloud code with the reverse engineer
280. 8:05anything framework reverse engineer the
281. 8:07reverse engineer anything repo to
282. 8:09basically help me figure out how it
283. 8:10works better so that I can use it better
284. 8:13and point my agents at it better and
285. 8:15instruct them better with it. So here is
286. 8:17what we have five main steps that I want
287. 8:19you guys to think about when you're
288. 8:20using this repo. Pick a feature route by
289. 8:23build summary first scene inferred
290. 8:25unknown and then rebuild and test. So we
291. 8:28already talked about this one a little
292. 8:28bit, right? not just pointing at the
293. 8:30whole app, but pointing at specific
294. 8:32behavior and functionality. Now, from
295. 8:34there, thinking about routing by build,
296. 8:36which means what the app is built with,
297. 8:38kind of picking the tools for that. So,
298. 8:40once you understand what sort of app or
299. 8:41software you're looking at and how it
300. 8:43actually exists, then you might be able
301. 8:45to change the way or alter the way that
302. 8:47the agent thinks about analyzing it to
303. 8:50use the right tools. And this isn't
304. 8:51something that I know off the top of my
305. 8:52head. I would just make sure I'm giving
306. 8:54the agent that context of you should be
307. 8:57thinking about this sort of stuff when
308. 8:58it comes to like the tech stack. Now
309. 8:59number three is about summary first. So
310. 9:01what happens is REA will basically
311. 9:03analyze the whole app and give you a bit
312. 9:04of a summary but the whole analysis for
313. 9:07Loom for example was over a gigabyte. So
314. 9:09basically the full analysis is turned
315. 9:11into a shorter summary for Claude to
316. 9:12actually fully read through. And it's
317. 9:14things like what it works, how this
318. 9:16thing does what it does and what it
319. 9:18couldn't yet figure out. And then what
320. 9:19happens is it kind of breaks that down
321. 9:21more granularly down to the modules and
322. 9:23to the traces. But you don't want to
323. 9:24have to rerun that entire summary, which
324. 9:26is something that you could easily get
325. 9:27caught in the loop of if you're not
326. 9:29being careful about how you instruct it
327. 9:31because this is already saved locally
328. 9:33within this project or somewhere in your
329. 9:35files. Now, this fourth one's really
330. 9:36important, which is once it starts to
331. 9:38dig through and understand what's going
332. 9:39on, Claude should be tagging these as
333. 9:42seen, inferred, or unknown. And these
334. 9:44are basically three different
335. 9:45categories. Meaning seen are things that
336. 9:48Claude actually found and were cited. So
337. 9:50for Loom 11 custom native add-ons, it
338. 9:53found it and it understands how they
339. 9:54work. Inferred is things that oh the
340. 9:57data and the code kind of points to this
341. 9:59being the way that it works, but this is
342. 10:01kind of more of a guess. We don't
343. 10:02actually know for a fact. And the reason
344. 10:04why this one's so important is because
345. 10:05if you accept some of these guesses and
346. 10:08assumptions that Claude might be
347. 10:09bringing back and then you build on top
348. 10:11of them as if they're fact, you can get
349. 10:13into some pretty messy situations. And
350. 10:16then of course we have things that are
351. 10:17completely unknown. Rea couldn't see it
352. 10:20inside of the native recorder or inside
353. 10:22of the the app that you're actually
354. 10:23using or looking at. And so it's these
355. 10:25types of things that you're able to do
356. 10:26with this reverse engineering anything
357. 10:28framework and repo that if you just
358. 10:30natively said, "Hey Claude, go look at
359. 10:32this app, research it, investigate it,
360. 10:34and then build it." That's where there's
361. 10:36a big difference on like how
362. 10:37professional and and I guess structured
363. 10:40this process feels when you use this
364. 10:41repo. And then of course, this is the
365. 10:43most important part of any build process
366. 10:45is the rebuilding and the testing. So
367. 10:47now you're taking all these findings
368. 10:48that you had grabbed and then you're
369. 10:51building on top of them. And you're not
370. 10:53just building, but you're also going
371. 10:54through and you're iterating and you're
372. 10:55actually testing these things out.
373. 10:57You're having it run these test cases.
374. 10:58You're having it actually process
375. 11:01things. So it's not just built, but it's
376. 11:02actually ran and tested and then
377. 11:04iterated on rather than just, you know,
378. 11:06shipping something once because it seems
379. 11:09like it's done. And the last things I
380. 11:10want to talk about here is that you
381. 11:11should never be copying exact code or
382. 11:13assets. There are definitely some legal
383. 11:15limits here when you are deeply
384. 11:16inspecting what's going on, but for the
385. 11:19most part, you also have to remember
386. 11:20that you're only looking at or cloud
387. 11:21code or whatever AI you're using is only
388. 11:23looking at things that actually exist
389. 11:25already and that you can grab. Meaning,
390. 11:27you know, you're not going to be telling
391. 11:28this thing to hack into like Loom's
392. 11:30server or to some app and software's
393. 11:32like private IP codebase. You probably
394. 11:34couldn't even get in there, but you
395. 11:36know, you're not going to tell your AI
396. 11:37to go hack into it. you're having it
397. 11:39analyze what it can already see in your
398. 11:41local files or if you have an API,
399. 11:43whatever it may be. It's just inspecting
400. 11:45how it already works and what already
401. 11:46lives there. And this whole process of
402. 11:48decompiling is something that people are
403. 11:50already doing because it's literally
404. 11:51there. It's public. It's it's in your
405. 11:52your hands. But humans obviously go
406. 11:54through it a lot slower than what an
407. 11:56agent would do. And also most of the
408. 11:58humans out there, I definitely couldn't
409. 12:00read through code and read through all
410. 12:01these traces and decompile it and then
411. 12:03build my own app based on that. The next
412. 12:05thing is that some apps require
413. 12:06different types of setup. you know,
414. 12:07based on, you know, we talked a little
415. 12:09bit earlier on where they actually live
416. 12:11and thinking about the tech stack and
417. 12:12thinking about how you need to set that
418. 12:14up. And then, of course, at the end of
419. 12:15the day, the results that you're going
420. 12:16to get from this thing are based on the
421. 12:18AI model that's actually writing the
422. 12:19code and inspecting the code. So, you
423. 12:21can't just assume that, oh, because you
424. 12:22gave, you know, your AI model this repo,
425. 12:26it's just going to build everything
426. 12:27magically. But what I do think is cool
427. 12:28about that is let's say you know you
428. 12:30have this software that you were
429. 12:31building with Opus 5.5 and maybe you
430. 12:33want to try with Sonnet 5.5 and it
431. 12:35wasn't possible earlier but now if
432. 12:37Sonnet 5.5 has all of this data that has
433. 12:39been done based on this investigation
434. 12:41maybe Sonnet 5.5 could build this for
435. 12:43you now and you could build this for a
436. 12:45lot cheaper than if you were running
437. 12:46Opus the whole time. So those are a
438. 12:47couple things that I wanted to talk
439. 12:48about. I will link all of this in my
440. 12:49free school community and I will drop
441. 12:51this repo in the description of this
442. 12:52video if you want to check it out. You
443. 12:54could also just Google search for it or
444. 12:55go on X and it will probably be the
445. 12:56first thing you see. So anyways, now
446. 12:58what I think is interesting about this
447. 12:59is a couple things. When I gave two
448. 13:02different coding agents the same prompt
449. 13:04except for one of them used the repo and
450. 13:05one of them didn't, the one that didn't
451. 13:07use the repo always finished first, but
452. 13:09the one that used the repo finished
453. 13:11slower but was much better results
454. 13:14because like I said, they were able to
455. 13:16use these frameworks inside of the REA
456. 13:18repo that understand how to break things
457. 13:20down a little bit better. It's it's this
458. 13:22whole REA investigation flow which is
459. 13:24what makes this so popular. I guess this
460. 13:27thing really is trending quick. It's got
461. 13:29192K stars. If I refresh, we're going to
462. 13:31go up to what 195 196K. And this was
463. 13:34probably 30 minutes ago that I last
464. 13:36refreshed this. Now, something else that
465. 13:37I think is really interesting is coding
466. 13:39agents are just going to get better at
467. 13:40this. Like I think that this whole idea
468. 13:42of reverse engineering now that this
469. 13:44framework has been blowing up on X like
470. 13:47I wouldn't be shocked if the next
471. 13:48version of Opus and Fable and Astra and
472. 13:50Soul and the other models have this
473. 13:52baked in a little bit better. But
474. 13:54anyways, really cool tool. It definitely
475. 13:56helps people that don't come from a
476. 13:57formal programming or software
477. 13:59engineering background build things that
478. 14:01feel a bit more professionally built.
479. 14:03And remember, the concept of reverse
480. 14:05engineering isn't anything new. You've
481. 14:06probably been doing it in your workflow
482. 14:07without even realizing it. we're able to
483. 14:09just basically give our agents a much
484. 14:12more specific PRD without having to go
485. 14:14through that painful process of building
486. 14:15a PRD because we already have a role
487. 14:18model that we're looking at of like this
488. 14:20is what we want now let's build it. So I
489. 14:22hope that that all made sense. I hope
490. 14:24that you can now sort of think about
491. 14:25reverse engineering in a different lens
492. 14:27which is it's not just about reverse
493. 14:29engineering software or apps. It's about
494. 14:31reverse engineering processes and skills
495. 14:34and things that are actually going on in
496. 14:35your day-to-day or in your business. go
497. 14:37out there, find someone who's done
498. 14:39something really well and proven that
499. 14:40they can do it well, and then understand
500. 14:42why that worked and how it it applies to
501. 14:45your context, to your personal
502. 14:46workflows. So, that was the video for
503. 14:49today. I hope you guys enjoyed, you
504. 14:50learned something new. I hope that all
505. 14:51made sense. If it did, please give it a
506. 14:52like. It helps me out a ton. And as
507. 14:54always, I appreciate you guys making it
508. 14:55to the end of the video, and I'll see
509. 14:56you on the next one. Thanks everyone.