DEV Community

LMSlay
LMSlay

Posted on Fully Autonomous

Reverse-engineer a creator's style: an AI persona and style guide from their latest videos

#ai #api #youtube #contentcreation

Disclosure: I build TubeText. This post uses our Video Breakdown AI Actor on Apify. The output below is from a run I did today.

When you study a creator, the useful question isn't "what is this video about?" but "what does this person do every time?": which hooks they open with, how fast they cut, what they sell, how they end. Answering that means watching a dozen videos with a notepad.

Here's how I get a first draft of that notepad from one API call: a persona and style guide built from a creator's latest videos, using Video Breakdown AI.

One call: a channel in, a style guide out 

curl -X POST "https://api.apify.com/v2/acts/tubetext~video-breakdown-ai/run-sync-get-dataset-items?token=$APIFY_TOKEN" \
  -H 'Content-Type: application/json' \
  -d '{"urls":["https://www.youtube.com/@mkbhd/shorts"],"maxVideosPerSource":5,"accountSummary":true}'
Enter fullscreen mode Exit fullscreen mode

The Actor takes the 5 newest Shorts, breaks each one down (hook, shot list, cut pacing, products shown, script), then condenses them into one extra row with "rowType": "account_summary". It works with a YouTube channel or an Instagram profile, and it needs at least 3 analyzed videos.

What came back for MKBHD's Shorts

The summary row, shortened:

{
  "rowType": "account_summary",
  "author": "Marques Brownlee",
  "videosAnalyzed": 5,
  "medianDurationSeconds": 66,
  "medianCutsPerMinute": 9.8,
  "medianShotCount": 11,
  "medianViews": 7170706,
  "accountSummary": {
    "content_pillars": [
      {"pillar": "Product Reviews & Deep Dives", "share_pct": 60},
      {"pillar": "Tech Industry News & Analysis", "share_pct": 20},
      {"pillar": "Concept & Speculative Tech", "share_pct": 20}
    ],
    "hook_patterns": [
      {"type": "visual surprise", "share_pct": 60, "example": "Hey Benny. Do a flip."},
      {"type": "bold claim", "share_pct": 40, "example": "All right, the folding iPhone is here. It's real. It's called the iPhone Duo."}
    ],
    "ctas": ["Get subscribed for more", "Watch the full video", "Follow for more"],
    "what_performs_best": "Videos featuring high-interest, 'impossible' tech concepts or viral-worthy hardware (like the iPhone Duo or Benny robot) significantly outperform standard industry analysis…"
  }
}
Enter fullscreen mode Exit fullscreen mode

It also returns a persona, tone and voice, on-screen text style, products and brands, posting cadence, and a style guide of 5 to 8 rules. Three of the seven it wrote for this channel:

  • "Start with an immediate visual or auditory hook that defies expectations."
  • "Prioritize 'show, don't tell' by using high-quality B-roll to demonstrate features."
  • "Keep on-screen text minimal, using it only for critical data points or emphasis."

Two things I find more useful than the prose:

  • The hook mix is countable. 3 of 5 Shorts open on a visual surprise and 2 on a bold spoken claim. Run it on three competitors and you can compare those shares directly.
  • The medians give you a brief. About 66 seconds, about 10 cuts a minute, 11 shots. That's a concrete target for an editor, not "make it punchy".

The per-video rows are still there

Each of the 5 videos also gets its own row, so you can check the summary against the evidence. For example, the hook row for "Apple's Big Price Hike":

{
  "text": "All right, so the memory price hikes have finally hit Apple and they have really hit.",
  "type": "bold claim",
  "delivered_as": "voice",
  "why_it_works": "It immediately addresses a major tech news story with a sense of urgency and direct impact on the viewer's wallet."
}
Enter fullscreen mode Exit fullscreen mode

Cost

Each video is $0.03 at Apify free-plan prices, plus $0.01 per minute past 3 minutes. The account summary row is free while it's in beta. This run was 5 Shorts, so about $0.15, and it took about 3½ minutes. Videos that can't be downloaded or analyzed aren't billed.

If a channel posts long videos, set "channelSort": "popular" to study its most-viewed uploads instead of the newest, and keep maxVideosPerSource small at first.

Caveats

  • 5 videos is a small sample. The shares (60%, 40%) are counts, not statistics. Use 10 to 20 videos for anything you'll act on.
  • The persona and style guide are a model's reading of the videos. Treat them as a first draft to check against the per-video rows, not as fact.
  • Instagram profiles often hide view counts. When that happens, the summary says so instead of guessing what performs best.

Questions or a platform you need? Open an issue on the Actor's page. I read every one.

Top comments (0)