2-Hour Buffett Interview → 44 Shorts (Data-Driven, Zero Gut Feeling)

How I reverse-engineered TikTok's top 100 Buffett clips to find what actually goes viral

Sarah Chen
Sarah Chen

SENIOR SOFTWARE ENGINEER AT TECHCORP....

Hi, it's Jennie.

I decided to use Warren Buffett interview clips for CashLingo's TikTok marketing. When you search "Warren Buffett" on TikTok, there are tons of videos with 100K+ views. That means there's massive traffic potential in this topic.

TikTok search results for Warren Buffett

But here's the thing.

When you cut shorts from a long-form interview, most people do it by gut feeling. "This part looks good, let me clip it." But that gut feeling is the difference between 877 views and 3.1 million views. Same person's interview, 3,500x difference.

So I decided to go with data, not gut feeling.


Step 1: Crawled the Top 100 Warren Buffett TikTok Videos

First, I had Claude Code crawl 20+ TikTok fan accounts using yt-dlp and collect the top 98 videos sorted by view count. Titles, views, creators — everything.

#1 was "Warren Buffett shares his best stock tip!" — 3.1M views.
#2 was "Warren Buffett on how much money is in his wallet!" — 1.4M views.
#3 was the McDonald's breakfast clip — 1.1M views.

That's just a list by itself. But when I gave it the right prompt, the "essence" came out.

Claude Code extracting viral essence from top 100 data

Step 2: The Essence = What Goes Viral vs. What Doesn't

When you look at the top 100 titles side by side, the pattern is obvious.

3.1M views title: "best stock tip" — specific, just one thing, ends with !
877 views title: "rules for making money" — abstract, multiple things, flat ending.

Same Buffett content. 3,500x difference. The formula that emerged:

[Billionaire] + [Everyday/Common subject] + [One specific thing] + [!] = Views

"best stock tip" = one tip → 3.1M.
"wallet" = one belonging → 1.4M.
"McDonald's breakfast" = one habit → 1.1M.

On the flip side, "rules for making money" with its broad scope and preachy tone? 877.


Step 3: Title First, Then Find the Clip

Just like YouTube — you design the thumbnail and title before you edit the video. For shorts cut from long-form, you need to decide the title of that short FIRST.

I downloaded the full transcript of CNBC's 2-hour Buffett interview (14M views on YouTube), and had Claude read through the entire thing with the essence as a filter: "What title could I post if I cut this section?"

44 clips confirmed in timeline table

Started with 80 candidates. Ran each one through the essence checklist:

  • Does it focus on ONE specific thing?
  • Is there a contrast gap or desire trigger?
  • Does the title match the S-tier pattern?
  • Is it NOT abstract/preachy/news-toned?

Deleted 36. Confirmed 44.


Step 4: Reverse-Engineered the Editing Format from Top 5

Clips are set. But how should they be edited? I didn't guess this either.

Downloaded the top 5 TikTok videos, extracted frames every 2 seconds, and analyzed each one's editing techniques. Additionally, I manually scraped ~80 comments per video to understand viewer reactions.

Top 5 analysis folder with videos and analysis files Detailed timeline analysis of #1 video Comment analysis with percentage breakdown

The conclusion was kind of shocking.

All 5 videos had:

  • BGM: None
  • Sound effects: None
  • Transitions: None
  • Motion graphics: None
  • Stickers: None

The only thing they did was crop the horizontal interview to 9:16 with the face zoomed in to 70-80%, and put a title text at the top. That's it.

Clean Authority format confirmed

The less editing, the higher the views. I named it the "Clean Authority" format.


Step 5: Auto-Produced All 44 Clips

With the format locked in, it was time to automate.

Built a Python script (produce.py) that does:

  1. ffmpeg to extract the clip segment
  2. Whisper for auto-synced captions
  3. OpenCV Haar Cascade for face detection → face-tracking crop
  4. ASS subtitle + top title overlay
  5. Final render at 1080x1920

Ran 3 clips in parallel, 12 batches total. All 44 videos completed in one session.

44 finished shorts in the output folder

The One Takeaway

When you cut shorts from long-form, clipping "what looks good" gets you 877 views. Clipping "what matches the 3.1M pattern title" gets you traffic.

You don't pull titles from content. You find content that matches viral titles. The order is reversed.

Data beats gut feeling. Every time.


Jennie

Responses (0)