Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

A designer with no coding experience spent 3 days and 700 revisions getting his AI-built video clipping extension to work

A motion designer who had never written code built a Chrome extension with an AI assistant. Two sliders on the YouTube timeline cut out the segment you need, up to 4K at 60 fps. The prototype was ready after about ten prompts, but making it work on other sites took three days and more than 700 iterations, meaning code fixes and restarts. The takeaway for practitioners: AI quickly produces a "everything works" scenario, while most of the time goes to bugs on real websites.

AI-processed from Habr AI; edited by Hamidun News
A designer with no coding experience spent 3 days and 700 revisions getting his AI-built video clipping extension to work
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1The extension sends the link and timestamps to a local script that runs yt-dlp and FFmpeg
Key metric #2If a download takes longer than 14 seconds, a fallback method kicks in
Key metric #3The GitHub repo has over 1,300 downloads, and auto-update is now disabled

Vibe coding is usually presented as writing one sentence and getting a finished app. The story of a motion designer who had never written code looks more honest. With an AI assistant, he built a Chrome extension that cuts a video fragment directly on the player's scrubber bar. A working YouTube prototype appeared after about ten prompts. But getting the extension to work on other sites took three days of testing and more than 700 iterations, meaning code fixes and restarts. That gap between "it worked once" and "it works for people" is what makes the story useful.

The author has worked with video for many years and knows UI/UX well. His daily frustration will be familiar to anyone who edits. Free download sites and Telegram bots either cap quality at 720p or make you watch ads. Desktop programs require a subscription and are inconvenient when you need a short clip rather than the whole video. For development, he used the Antigravity environment, mostly with the Argon model.

How it works: two parts instead of one

The idea is simple: you watch a video in your own Chrome, switch the mode, and two sliders appear over the scrubber bar. You drag them along the live video, see the actual frames, and grab the fragment at the resolution you want, up to 4K at 60 frames per second.

The AI immediately pointed out a limitation. The extension can see the page and the player, but it cannot merge video and audio: on YouTube, everything above 720p comes as separate tracks. It also cannot quickly cut heavy videos frame by frame. Free command-line utilities do this work, yt-dlp (downloading) and FFmpeg (merging and cutting), but they cannot see the open browser tab.

So the app was split into two parts:

  • the extension finds the video on the page, draws a window with quality buttons, and attaches the sliders to the timeline;
  • a local helper script receives the link, quality, and timestamps, runs yt-dlp and FFmpeg, and returns the finished file to the browser's standard save dialog;
  • the install.bat installer downloads the components from official sites and connects them to Chromium-based browsers.

The first run produced a 1080p clip with sound a couple of seconds after the sliders were moved.

Three days instead of "piece of cake"

The author asked whether downloading from all sites was possible. The AI replied that it was easy and everything was ready. Then came three days of debugging across dozens of platforms: YouTube, VK Video and VK Clips, RuTube, Dzen, Odnoklassniki, TikTok, Instagram Reels, Twitch, Reddit, Niconico, and Bilibili. Typical problems included:

  • the player was hidden in closed page blocks or embedded in an iframe from another domain, and the extension missed it. The app was taught to look inside such containers;
  • in endless feeds such as Shorts, VK Clips, and TikTok, five to ten videos are loaded at once, and the extension grabbed a neighboring one. Now it identifies the video in the center of the screen and outlines it with a green frame;
  • in VK Clips, the program started downloading a two-hour episode instead of the short clip linked in the description;
  • on vertical and widescreen videos, the sliders drifted away from the timeline, so coordinate binding had to be adjusted for different players;
  • with a VPN or the Zapret program, the browser plays video normally, but local downloading hangs: in one test, a cut from Shorts stalled for five minutes. A DNS workaround and a watchdog timer were added. If a download takes longer than 14 seconds, the app switches to a backup multithreaded method.

For practitioners, this is the key lesson. AI quickly assembles a near-perfect scenario, and the rest of the work consists of edge cases. They have to be found by hand and described precisely to the model. The author says he spent entire days hunting bugs. The article says nothing about token usage or costs, so the economics of this kind of development cannot be assessed from it.

Backlash, security, and interface

After the GitHub release, some commenters attacked "vibe coders." Others looked at the manifest, saw install.bat and access to cookies, and suspected password theft. According to the author, the project contains five plain-text files, none of his .exe files, and no data collection. Cookies are needed locally so yt-dlp can download 18+ videos and clips from private groups. The temporary file is deleted after download.

One criticism turned out to be fair: silent auto-updating of scripts from GitHub once a day. If the author's account were compromised, malicious code would immediately reach all users. The author added an Auto-update toggle. It is a good example of why AI-built projects should be checked for such supply chain risks before publication.

Then features piled up: adjusting boundaries to the second and frame, a lock on fragment length (handy for 15-second inserts), selecting an episode by the uploader's timestamps, and a bug-report button. The interface grew bloated, so version 1.2.0 simplified it. By default it keeps the Full Video / Cut Fragment switch, a resolution grid, and a Download button, while everything else is moved under Advanced View.

The author admits the app does not work flawlessly on every site, since there are simply too many players. At the time of publication, the project had more than 1,300 downloads on GitHub. For founders and product managers, the conclusion is straightforward: a non-technical person with a good grasp of the task and interface can bring a useful tool to a working state, but testing and security remain their responsibility.

Story

Key Questions Answered

How it works: two parts instead of one?

The idea is simple: you watch a video in your own Chrome, switch the mode, and two sliders appear over the scrubber bar. You drag them along the live video, see the actual frames, and grab the fragment at the resolution you want, up to 4K at 60 frames per second. The AI immediately pointed out a limitation. The extension can see the page and the player, but it cannot merge video and audio: on YouTube, everything above 720p comes as separate tracks. It also cannot quickly cut heavy videos frame by frame.

What are the key details on Three days instead of "piece of cake"?

The author asked whether downloading from all sites was possible. The AI replied that it was easy and everything was ready. Then came three days of debugging across dozens of platforms: YouTube, VK Video and VK Clips, RuTube, Dzen, Odnoklassniki, TikTok, Instagram Reels, Twitch, Reddit, Niconico, and Bilibili. Typical problems included: the player was hidden in closed page blocks or embedded in an iframe from another domain, and the extension missed it.

What are the key details on Backlash, security, and interface?

After the GitHub release, some commenters attacked "vibe coders." Others looked at the manifest, saw install.bat and access to cookies, and suspected password theft. According to the author, the project contains five plain-text files, none of his .exe files, and no data collection. Cookies are needed locally so yt-dlp can download 18+ videos and clips from private groups. The temporary file is deleted after download. One criticism turned out to be fair: silent auto-updating of scripts from GitHub once a day.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…