How to run your first AI UGC campaign (step-by-step guide) — Transcript

Full transcript

1. 0:00What the [ __ ] is UGC creators? Well,
2. 0:02when you see a video like this on
3. 0:03Instagram or Tik Tok that you saw it is
4. 0:06a real user generally share a product
5. 0:08they like is most likely a UGC content
6. 0:11created by the product company. They
7. 0:12typically pay $ 20 to $50 per each video
8. 0:14and sometimes even thousands of dollar
9. 0:16for just one content. And here comes
10. 0:18Jamaica quickly explain how it works in
11. 0:1920 seconds. I did some research and
12. 0:21found examples of videos these campaigns
13. 0:23produce. They were extremely low effort
14. 0:26but I could see why they worked.
15. 0:31The basic UGC strategy is pretty simple.
16. 0:34Find a bunch of kids with too much time
17. 0:36on their hands. Have them create new
18. 0:38social media accounts dedicated to
19. 0:40posting about your app and pay them $ 20
20. 0:42to $50 per video. You're not paying for
21. 0:45their audience because they usually
22. 0:47don't have one. And you're not posting
23. 0:49these videos on your own account because
24. 0:51the whole point is for them to look like
25. 0:53normal people who just happen to
26. 0:55discover your app. And since I already
27. 0:57know what CDES 2.5 type of model is
28. 1:00capable of producing, it's only
29. 1:01reasonable for me to think there must be
30. 1:03a way for us to produce those
31. 1:04hyperrealistic UGC content. But for the
32. 1:07past few days, I tried all sort of
33. 1:08different ways. And I think I finally
34. 1:10cracked it. A workflow that allow you to
35. 1:12produce hyperrealistic UGC content.
36. 1:14They're going to drive down the cost
37. 1:15from 20 to $50 to only 20 cents per
38. 1:18second. And all those UGC content will
39. 1:20be produced by cloud code or your codeex
40. 1:22directly by utilizing free plug-in like
41. 1:24track which is open source project for
42. 1:25your agent to access also different data
43. 1:27and tooling. I'll take you through step
44. 1:29by step so you guarantee can produce the
45. 1:31same result as I get. So one thing I
46. 1:33always started with is try to get a
47. 1:35bunch of videos that is working really
48. 1:36well in our vertical and analyze them
49. 1:39because the crux of the UGC content is
50. 1:41literally two parts. One is a hook which
51. 1:44probably the most important thing and
52. 1:46second part is content demo that
53. 1:48showcasing your product and typically
54. 1:50the way to scale it is the same product
55. 1:52demo can be paired with five or 10
56. 1:54different hooks. So you can test out
57. 1:56huge amount of variations to see what
58. 1:58actually works and by just watching the
59. 2:00trending content in your vertical
60. 2:02recently you get understanding about how
61. 2:05you should design those hooks and also
62. 2:07help you find what type of face or
63. 2:09characters actually resonate with your
64. 2:11audience because believe it or not with
65. 2:13the same script change your person might
66. 2:16significantly change the impression.
67. 2:17We're going to use this free tool called
68. 2:19track which full contact actually uses
69. 2:21because when my team were running
70. 2:22different loops to drive go to market
71. 2:24one of the biggest problem I have was
72. 2:26that agent needs access to a lot of
73. 2:28different data as well as tools but most
74. 2:30of those data I locked behind some pay
75. 2:32wall that cost me a few hundred no
76. 2:34matter how many times I actually need
77. 2:35that data that's why I bu so that you
78. 2:38only need to connect one key and have
79. 2:39access to all sort different service and
80. 2:42don't need to pay for subscription but
81. 2:44pay for usage only and because I was
82. 2:46using this for my own company. It has
83. 2:48really rich data for your agent to find
84. 2:50what's actually trending and working in
85. 2:53your social channels or from your
86. 2:55competitors. All you need to just sign
87. 2:56in, choose the agent that you're using
88. 2:58and copy this prompt to your agent. It
89. 3:00will be able to just set up. After it's
90. 3:02done, just give a prompt. Help me pull
91. 3:04recently trending videos in our
92. 3:06verticals AI tooling. Within one minute,
93. 3:08it just got me this list of videos that
94. 3:10is actually trending recently on Tik Tok
95. 3:13and Instagram for my specific vertical.
96. 3:15and I can look through the hooks, find a
97. 3:17character that we like and create a AI
98. 3:19character that get that vibe, but it's a
99. 3:21different person. Among all those
100. 3:23videos, I quite like this character look
101. 3:26smart, approachable. If she talk about a
102. 3:28solution where she makes money, I
103. 3:30probably will believe it. And our goal
104. 3:31is take a character that actually exists
105. 3:34and we like, grab his vibe and generate
106. 3:36our own version so that we can create
107. 3:38account based on that character and make
108. 3:40a UGC content out of it. And this to be
109. 3:43honest is really really difficult. For
110. 3:46context, if all you want to do is just
111. 3:48generate a realistic video, that is
112. 3:50easy. This is a video I got just one
113. 3:53shot from Cance 2.5 about a girl talking
114. 3:56about a script. And this one looks
115. 3:58absolutely real. But the problem is that
116. 4:00I have no control about how the person
117. 4:02going to look like. And if I do have a
118. 4:04picture of a person already and want to
119. 4:07make a realist video based on that, this
120. 4:09is also easy. This is what we got. To
121. 4:11me, this already super realistic. If I
122. 4:14show to my friends, they probably won't
123. 4:15be able to tell that this is AI
124. 4:17generated. But the real hard part is
125. 4:18that you want to generate videos with a
126. 4:21character that has certain vibe that you
127. 4:23really like. I actually try a few
128. 4:24different approach. For example, I was
129. 4:26trying to create a AI character based on
130. 4:29this lady who were making a lot of good
131. 4:31videos from Google Cloud YouTube channel
132. 4:33and take a screenshot of her and feed to
133. 4:35GP2 image 2.5 and Gemini 3. The result I
134. 4:39got is something like this. The left one
135. 4:41is GP image 2.5 and right side is Gemini
136. 4:443 Pro. Both are the state-of-the-art
137. 4:47models, but they just didn't look like a
138. 4:49real person. Something about them feels
139. 4:51really off and I can't really pinpoint
140. 4:53why. But luckily, I got an amazing
141. 4:55prompt from a friend Harry which totally
142. 4:57changed output. This result I got with
143. 4:59the right prompt from him. You can see
144. 5:01that it actually generated a character
145. 5:04that really got a vibe of the original
146. 5:06person and it looks real. It didn't look
147. 5:08like those normal AI slop image and the
148. 5:11key thing here is that you want to
149. 5:12restrict the exploratory space for the
150. 5:16model because at a default even batch
151. 5:18model like GV image 2.5 still have some
152. 5:21out ofbox characteristic like for the
153. 5:23eyes is always quite big and face often
154. 5:26it just got those type of doll face and
155. 5:28large model often include quality
156. 5:30booster in the prompt like 4K AK ultra
157. 5:34detail which will often lead to those
158. 5:35high clarity HDR type of picture then
159. 5:38make it looks like very AI slop and the
160. 5:40way we solve that is output a JSON based
161. 5:44prompt like this they almost force model
162. 5:46to define every single details they
163. 5:48otherwise will omit and revert to
164. 5:50default and I tried this skew a few
165. 5:52times it g me really amazing results
166. 5:54like I'm feeding this pretty famous
167. 5:57Chinese tech vlogger and is able to
168. 5:59output a new character that look
169. 6:02different but got original vibe same
170. 6:04thing of another Chinese AI podcaster
171. 6:06and generating new character with the
172. 6:08same vibe. I even got this new character
173. 6:10by feeding image of Jen Yan from Silicon
174. 6:13Valley and I think it did a pretty good
175. 6:15job. Most importantly, this looks real.
176. 6:17It didn't look like those AI slop images
177. 6:20and he was kind enough to turn that
178. 6:21knowledge into a reusable skill that
179. 6:24allow you to feed model any screenshot
180. 6:26and character and I'll put a new
181. 6:28character that looks different but got
182. 6:31the vibe. And I have put the link to his
183. 6:32skills in the description below so you
184. 6:34can check out for free. And I highly
185. 6:36recommend you go and follow him on
186. 6:37Twitter. He's one of the best builder I
187. 6:39met in my life. So this is the first
188. 6:41trick. You want to have Jason B prompt
189. 6:43like this that define every single
190. 6:44details to constrain models explorative
191. 6:47space. And second trick I found is that
192. 6:49even you got a prompt right different
193. 6:50models still have different performance.
194. 6:52Here I use same prompt across five
195. 6:54different models. And please comment
196. 6:56below. Let me know which one you feel is
197. 6:58best. But based on my test, I actually
198. 7:00found Gemini 3 Pro leads to the best
199. 7:02results. Like technically JP2 image 2.5
200. 7:06should be the best one. The character
201. 7:07generated somehow has this patterns
202. 7:09where the eyes didn't look as real and
203. 7:12the face generated more like those
204. 7:13dollface versus Gemini 3 pros results
205. 7:16are super realistic and we actually add
206. 7:18all the major image and video generation
207. 7:21model on track as well. So with a track
208. 7:23skew you are able to get your cloud code
209. 7:25or codecs to generate character with
210. 7:28super cheap price because we don't take
211. 7:30any markup. So you can actually do
212. 7:31things like help me generate character
213. 7:33using portrait clone skew with GPT image
214. 7:362.5 Gemini 3 pro Cream so I can compare
215. 7:40based on this screenshot of the original
216. 7:41person and then it is able to call
217. 7:43different models so you can easily
218. 7:45compare the results and choose the best
219. 7:47one you want. And next now we get
220. 7:49character ready we want to create the
221. 7:51actual UGC content. All you need to do
222. 7:53just feed to the CDS 2.5 model and use
223. 7:57this as an image reference. it will do a
224. 8:00really good job. The only problem is
225. 8:01that as a default the C dance 2.5 model
226. 8:05will refuse to take hyper realistic
227. 8:07image like this to prevent you from
228. 8:09impersonate someone and making video out
229. 8:11of it and platform like Hickview they do
230. 8:13provide end point that allow you to
231. 8:15bypass that but they also charge a huge
232. 8:17amount of margin they're going to make
233. 8:18your cost at least double or triple and
234. 8:20the correct way to do that is you buy
235. 8:22this API from by plus directly there's
236. 8:25some advanced creation rights and even
237. 8:27with free version that has this premium
238. 8:29creation benefit package Page entry will
239. 8:31allow you to upload image and create a
240. 8:33digital character based on that and then
241. 8:35it can generate video based on this
242. 8:37character that you register but it does
243. 8:39take a few different hurdles to get
244. 8:41approval from it. However, on track we
245. 8:42actually have this CS 2.5 restriction
246. 8:45and point or automated process to take a
247. 8:48picture upload as a assets for the
248. 8:50character bypass the whole restriction
249. 8:52and we're going to offer the cheapest
250. 8:53possible price on the market and take 0%
251. 8:55markup. So you can use a cdance 2.5
252. 8:57endpoint just like you buy directly from
253. 9:00by plus and all you need to do just to
254. 9:02track your cdance 2.5 less restriction
255. 9:04model to generate video using script
256. 9:07with this character in generally it can
257. 9:09make a really good job based on just one
258. 9:11image but this is one problem you will
259. 9:12probably notice voice it generate at
260. 9:14default didn't sounds right and trick I
261. 9:16found here is that you can also get
262. 9:18audio clip from original video that you
263. 9:20like then pass this audio as audio
264. 9:22reference this what I found is shortest
265. 9:24cut to get the whole video rhythm and uh
266. 9:27speech feels right because there are so
267. 9:29many details it's hard to otherwise
268. 9:31describe in text and this is final video
269. 9:33I got
270. 9:33>> 3:00 a.m. Last night, my agent was
271. 9:35researching hundreds of customers and
272. 9:37emailing each of them to book meetings
273. 9:38for me. I woke up to 12 replies and four
274. 9:41meetings on my calendar. Here's exactly
275. 9:43how I set it up.
276. 9:44>> And the only last part is a demo
277. 9:45content. And to just cut a cost, I found
278. 9:48in the demo content, you don't really
279. 9:50have the character showing. You
280. 9:52literally can just you record a video
281. 9:54and then use 11 lab official audio to
282. 9:56clone the voice and output the audio
283. 9:58clip and merge together with demo video
284. 10:00you make. Here I just manually use my
285. 10:02phone to record some demos of how people
286. 10:04can use plugin to get a codec doing GTM
287. 10:07work like find the lead doing the
288. 10:09autoreach. Then I create a voice on fish
289. 10:10audio using the original audio clips
290. 10:12that I got from the original video and
291. 10:14put together a product demo clips that
292. 10:16has real voice over and captions. And
293. 10:18once I get demo clips right, I can just
294. 10:20pair it with five or 10 different hooks
295. 10:23AB testing them to see which one
296. 10:25actually works. And here is how I did
297. 10:27it. I basically give track to my cloud
298. 10:29code and ask it to pull recently
299. 10:31trending videos in my vertical and I
300. 10:33give it some reference account that I
301. 10:35think is doing pretty good job like
302. 10:36composio then analyze and figure out the
303. 10:38hooks that I can be using for this demo
304. 10:40and grab a bunch of different trending
305. 10:42videos that's getting a lot of views and
306. 10:44because model nowadays are pretty good
307. 10:46at analyzing audios and transcript is
308. 10:49able to just extract hooks they were
309. 10:50using like this one video from composer
310. 10:52got 3.3 million impression they I email
311. 10:55over 5 million people and asked each for
312. 10:57$1. Okay, interesting. I assume they are
313. 11:00using this video to showcase like Gmail
314. 11:03integration I guess. So those
315. 11:04information are really useful for
316. 11:05yourself and agent to analyze a pattern
317. 11:07which is have pretty stunning claim and
318. 11:10showcase some numbers pair with results.
319. 11:12They asked to brainstorm some options
320. 11:14and one I actually pretty like 3:00 a.m.
321. 11:16last night my agent was researching
322. 11:18hundreds of customers and emailing each
323. 11:20one of them to book a meeting for me. I
324. 11:22woke up to 12 replies and four meetings
325. 11:24on my calendar. Here is exactly how I
326. 11:27set up. So I'm going to give it a prompt
327. 11:29that I like one now. Please use C dance
328. 11:312.5 with character image that we created
329. 11:34before and audio reference to generate a
330. 11:36video for me. And I actually created
331. 11:38this UGC talking head video skill that
332. 11:41is able to take in the request and
333. 11:43design the script and duration grab a
334. 11:45voice reference generate the prompt for
335. 11:47the video and even adding the caption
336. 11:49and headers to make a final video. If
337. 11:51you're interested this skill I put in
338. 11:52our GitHub repo of track already so you
339. 11:54can go and grab it. All I need is just
340. 11:56send this request then it will pull all
341. 11:58the resource and use a se dance 2.5 less
342. 12:01restriction model that we have on track.
343. 12:04And after a few minutes it is able to
344. 12:05generate full UGC clip for only $2.67.
345. 12:10And here's the result. 3:00 a.m. last
346. 12:12night my agent was researching hundreds
347. 12:13of customers and emailing each of them
348. 12:15to book meetings for me. I woke up to 12
349. 12:17replies and four meetings on my
350. 12:19calendar. Here's exactly how I set it
351. 12:21up. First, go to chat GBT. Search for
352. 12:23the track plugin. Then just tell it who
353. 12:26you're looking for. It researches every
354. 12:28lead, writes each one a personalized
355. 12:30email, and sends it from my Gmail.
356. 12:32>> So, this is pretty much the whole
357. 12:33workflow from find trending content and
358. 12:35character to create a AI character that
359. 12:38going to resonate with your audience to
360. 12:39actually generate those UGC content at
361. 12:42scale. And if the video you're
362. 12:43generating is a bit more complicated
363. 12:45than those type of software demo video,
364. 12:48then typical flow is that you want to do
365. 12:50a more detailed character design which
366. 12:52is not just front-facing but different
367. 12:55angles and then use image generation
368. 12:57model to create a storyboard to align
369. 12:59things then feed those storyboard plus
370. 13:02character to Cance 2.5 and generate the
371. 13:04final video. But for most of software
372. 13:06you see is typically pretty simple. I
373. 13:09found with one image is already good
374. 13:11enough. So this is pretty much the whole
375. 13:13flow and as mentioned before we be spend
376. 13:15a lot of time building track which is
377. 13:17like open router for data and agent
378. 13:19toolings. So you can just connect to one
379. 13:20service and will give you access to more
380. 13:22than 3,000 different premium data and
381. 13:24tools across people and company
382. 13:26enrichment buying signal SEO and AEO and
383. 13:29social. It even have free endpoint for
384. 13:31you to manage your social and ads
385. 13:33accounts through OOS for free. So you
386. 13:35don't need to pay additional service for
387. 13:36like buffer potato or uni. And it's
388. 13:39fully open source, so you can host
389. 13:40yourself or come and contribute. And
390. 13:42please comment below if you have any
391. 13:43questions or feedback.