The bait, then the rug-pull.
A clean 12-second open: state the deliverable, show the tool already running with 20M-view scrape results, promise a three-part structure. Then the entire rest of the video is just executing that promise without drift.
What the video promised.
stated at 00:02“I'll show you how to use Code to create viral TikTok videos that will generate your sales afterwards.”delivered at 08:23
Where the time goes.

01 · Cold open + proof
States the deliverable, shows the tool already scraping 20M-view competitor videos, names the outcome — 10x viral chances.

02 · Three-part roadmap
Slide deck declares the three parts of the video. Oleg restates it twice — signals the structure clearly upfront.

03 · Self-intro + Authority AI
5-years-in-AI credibility play, mentions his company Authority AI runs 90% automated using these same systems.

04 · Part 1 — Creators tab
UI tour: six beauty/lifestyle creators (Emira, Alix Earle, Lydia Millen, etc.), live demo adding Kendall Jenner by URL, scraping confirmation.

05 · Part 1 — Configs tab
Defines a 'Bella Glow' config: brand description, Gemini analysis prompt (hook/retention/CTA), Claude output prompt (three script variants per concept).

06 · Part 1 — Run Pipeline + Videos
Pipeline params: 20 recent videos per creator, top 3 selected, last 30 days only — to skip stale trends. Output is per-video analysis + tailored concepts.

07 · Part 2 — Install VS Code + Claude Code
Download VS Code, open the project folder he shares, install Claude Code extension. Notes Claude Code requires paid tier ($19/mo minimum).

08 · Part 2 — API keys (Apify / Gemini / Anthropic)
Walks through the .env file. Apify for scraping, Gemini for multimodal video analysis, Anthropic API for concept generation. Each is a separate console + key copy.

09 · Part 2 — npm run dev
Run the dev server, open localhost in browser, app is live. Notes you don't need Claude Code itself to use the app — only to modify it.

10 · Part 3 — Feed brand context
Tip 1: have Claude scrape your own website / LinkedIn / Instagram via Apify so it learns your ICP, products, and funnel before generating concepts.

11 · Part 3 — Declare filming constraints
Tip 2: tell the system what you can't film (solo creator, fixed location, no group shots) so it stops generating un-replicable ideas.

12 · Part 3 — Mine competitor comments
Tip 3: scrape thousands of TikTok comments on viral videos to find gaps / unanswered ICP questions you can address in your own content.

13 · Part 3 — Ingest your own history + CTA
Tip 4: scrape your own TikTok with Apify, learn what worked, feed the patterns back in. End-screen CTA points to a second Claude-Code-marketing video.
Visual structure at a glance.
Named ideas worth stealing.
3-API Viral Concept Pipeline
- Apify — scrape competitor TikToks
- Gemini — analyze hook/retention/CTA visually + audibly
- Claude (Anthropic API) — generate tailored scripts in your defined output shape
Each API does one job. Apify gets the raw data, Gemini is the only one that 'watches' the video, Claude is the writer. Clean separation of concerns.
Pipeline Filtering Defaults
- 20 most recent videos per creator
- Top 3 by virality
- Past 30 days only
Recency filter is the underrated piece — kills stale trends before Claude even sees them.
AI Context Stack for Personalization
- Brand context (site, LinkedIn, IG via Apify)
- Filming constraints (what you can/can't shoot)
- Audience signal (competitor comments)
- Your own performance history (your TikTok scraped)
Four context layers that turn generic AI ideation into something actually executable for your specific situation.
Lines you could clip.
“Our chances of going viral and generating sales will be 10x compared to randomly coming up with ideas or asking ChatGPT, hey, what should I film?”
“Even though this tool was built using Cloud Code, don't worry. It is very simple. Even if you have zero experience with it, I'll show you how to set it all up.”
“If you're not specifying that, you will get ideas that you just feel like, that's not for me. I'm wasting my time. At the end of the day, I will say screw the system. I'll do it all manually. And that's not what we want.”
“I cannot randomly make a video about fitness and how to grow huge biceps. Obviously, it's not specifically for you.”
How they spent the runtime.
Things they pointed at.
How they asked for the click.
“Check out the guide in description down below. And if you're interested, here's another video about how to use Cloud Code to solve different marketing tasks, like content creation, cold outreach, and even ads creation. Check it out.”
Soft. The real CTA is the description link to the project folder, which gates Claude Code Pro signups + the three API accounts. The end-screen video pitch is bolted on. He never explicitly says 'subscribe' — the lead magnet does the conversion work.
Word for word.
Steal the lead-magnet-tutorial format.
The video IS the funnel: the deliverable is a working tool, gated behind a paid AI subscription, taught with such clarity that the viewer feels they 'earned' it.
- Open by showing the tool already running on real data — not a slide, not a face-cam promise. Proof first, story second.
- Declare the three-part structure on a slide at 00:42, then literally return to that slide three more times. Free re-engagement points without saying 'stay tuned'.
- Stack three APIs by job — scrape / analyze / write — and name which AI does which. Viewers love clear separation of concerns more than they love magic.
- Build the lead magnet so it requires the paid product (Claude Code Pro $19/mo) to run. The tutorial is the funnel.
- Put the personalization layers (brand context, constraints, audience signal, your history) in Part 3, not Part 1 — so the value compounds even after viewers stop watching.
- Don't pitch the next video until the last 20 seconds. The whole video is the pitch; the end-screen is just plumbing.
What this could mean for you.
The system works because it strips your TikTok ideation down to a repeatable loop — copy what's already winning, filtered through what you can actually shoot.
- Pick 5-10 creators in your niche who consistently hit big numbers. Don't aim broad — make them as close to your situation as you can.
- Before you scrape anything, write down your filming constraints in one paragraph: who's in the video, where you shoot, what you won't do on camera. This single doc kills 80% of unusable AI suggestions.
- Don't just look at hooks. Read the top 50 comments on each viral video. The questions people ask are the next 10 videos you should make.
- Set a 30-day recency filter on whatever you analyze — old viral patterns are dead patterns.
- If you already post, treat your own back-catalogue as research. The audience you already have is the most honest data you'll ever get.
- Don't run this once and ship 20 videos. Run it weekly. Trends decay, and the system only stays sharp if your input data does.




































































