In today’s world, where content reigns supreme, video uniqueness becomes not just desirable, but critically important. Whether it’s to bypass strict anti-plagiarism algorithms, <a href="/ru/blog/massovyy-posting-video-kak-avtomatizirovat-distributsiyu-kontenta-na-10-akkauntov/" title="Массовый постинг видео: как automate content distribution to 10+ accounts?">scale advertising campaigns, or adapt for various platforms, the ability to transform one video into hundreds of unique variations opens up new horizons. This article will explain how we approached this task, using innovative methods and automation, not only to avoid bans but also to make valuable content accessible to a wide audience.
Introduction: Content Repurposing to Bypass Anti-Plagiarism
Content repurposing is not just repackaging. It’s a hacker’s approach to adapting videos for different platforms and purposes, especially when it comes to bypassing ad bans or scaling creatives. Andrey Karpathy, frustrated with YouTube, created HackerTube over a weekend – a showcase of 6454 videos that the Hacker News community had curated over the years. This idea prompted us to create our own fork, but with a unique feature: we adapted these videos for a Russian-speaking audience, turning them into brief textual reviews.
Creating a “Video Matryoshka”: From 6000 to 30 Interviews
We took Karpathy’s original data, published under the open ODC-By license, and focused on interviews and podcasts, which is our profile in Signal magazine. From 6454 videos, we selected 435 interviews, then filtered out news “one-offs” and unmentioned videos, leaving 310. Finally, 30 of the most relevant interviews were manually selected.
- Original database: 6454 videos.
- Filtered by type: 435 interviews/podcasts.
- Selected after cleaning: 310 videos.
- Manual editor selection: 30 final interviews.
This process is an example of how to “glue” and “split” a large amount of data to obtain targeted content. We used editorial selection, focusing on technology and business, ignoring medicine and politics, even if they had a high rating.
Uniqueness Algorithm: How We “Repackaged” Videos
Our pipeline runs locally on Python, without databases, which allows for quick generation of variants for A/B tests. The process consists of several steps:
- Subtitle Extraction:
yt-dlpis used to obtain automatic subtitles. - Block Segmentation: Subtitles are divided into blocks approximately one minute long, each with a timestamp.
- Card Generation: Claude Sonnet analyzes blocks and comments from Hacker News, returning JSON with key points, quotes, and a verdict.
- Verification of Authenticity: The code cross-references quotes with the original transcript.
- Static Page Assembly: Python generates HTML, and Nginx serves the content.
We also create unique covers by coloring the middle frame of the video in two colors. This is part of the uniqueness package, which allows for adaptation for TikTok, Shorts, Reels, without resorting to external sources.
Quote Verification: Battling the Model’s “Video DNA”
The main rule of our pipeline: the model cannot simply claim that the hero said something. The original English text of each quote is searched in the transcript. If an exact match is found, the quote receives the status “exact.” If a similar fragment is found, it’s “almost.” Otherwise, the quote is deleted.
“Out of 30 quotes, 26 matched exactly, 4 almost, and none had to be discarded.”
This helps combat the “video DNA” that the model might “invent.” Timestamps work similarly, ensuring that the link leads precisely to the desired moment.
Model Errors: Where the “Matrix” Fails
During the review of 30 cards, interesting errors were identified:
- Invented Facts: The model attributed bankruptcy to Chris Sacca, although he refused it to preserve his reputation.
- Time Errors: Karpatiy’s conversation with Lex Fridman was dated “when ChatGPT just appeared,” although the video was released a month before that.
- Unverifiable Claims: Phrases like “the only long interview” or “this is not in any textbook” without evidence.
These examples show that deceiving the algorithm regarding facts is harder than it seems. The model is prone to “smooth generalizations,” which often turn out to be fabrications.

“Don’t watch”: why the model cannot give harsh verdicts
We aimed for each card to contain a verdict: watch in full, watch a segment, or the card is enough. However, the model proved reluctant to make such decisions:
- First pass: The model did not advise «not to watch» long interviews, but offered one-hour segments.
- Prompt tightening: The requirement for a segment not exceeding 25 minutes and an increase in the proportion of «card is enough» led to the disappearance of the latter and the relocation of one-minute videos to «in full».
As a result, four «card is enough» verdicts (for Bill Gates, Geoffrey Hinton, George Hotz, and Chris Sacca) were manually assigned. «It seems that telling a great person ‘you don’t have to watch’ remains an editorial job.»
Relevance and timeless ideas: is it worth «creating new content»?
Some videos in our selection are up to ten years old or more. Although this initially raised questions, it turned out that great ideas and thoughts do not age. Jim Simons’ interview about mathematics, Steve Jobs in 1994 about creating the world, Richard Feynman in 1983 about physics — all of them are still relevant today.
«Old stuff that hasn’t aged»
- Jim Simons on creating a successful fund.
- Steve Jobs (1994) on the world being created by people no smarter than us.
- Richard Feynman (1983) on physics without formulas and a blackboard.
This confirms the idea that video uniqueness does not always mean creating completely new content. Sometimes it is enough to adapt for different social networks and update the presentation so that «old» material plays in a new way.
Conclusion: content scaling and the future of HackerTube
Our project demonstrates how to scale content, transforming thousands of hours of video into short, informative cards in Russian. This not only allows you to bypass anti-plagiarism and uniquify to avoid bans, but also makes valuable knowledge accessible to a wide audience. We continue to work on the «Signal» magazine, focusing on fresh interviews, but this experiment showed the potential of content multiplication. What to do next with our «Tube»? Your advice is very important. Leave them in the comments or write to us. And you can get a behind-the-scenes look at all of this in our Telegram channel.
Frequently Asked Questions
What is content repurposing in simple terms?
Content repurposing is the process of re-adapting existing content into new formats or for a new audience. For example, turning a webinar into 50 short clips for TikTok or Reels.
How to make 100 unique videos from one?
To create many unique videos from one, you can use various methods: changing the background, voice timbre, adding graphics, cutting into short fragments, changing the order of information presentation. Using AI models allows you to automate this process, creating creative variations for A/B tests.
Should I order or shoot a new video?
It depends on your goals and budget. A price comparison from scratch shows that repurposing video for advertising or repurposing it is often cheaper than creating new content. If the original material contains valuable ideas, adapting it can be more effective.
What is “video DNA” and how to trick it?
“Video DNA” is a metaphor for the unique characteristics of a video that anti-plagiarism systems use for identification. You can trick anti-detection by changing visual and audio parameters, narrative structure, and by using various uniqueization methods to bypass bans.
