I Built an AI Tool That Extracts Tech Stacks from YouTube Videos

Every developer has done this: you open a 45-minute "Build a SaaS in a weekend" video, and by minute 12 you're already pausing every 30 seconds to screenshot a terminal command or squint at a VS Code sidebar to figure out which ORM the creator just installed. You rewind three times because you missed the name of the vector database. By the time the video ends, you've spent an hour to extract maybe eight lines of useful information: a framework name, two library names, one hosting provider. I built VidScope because I got tired of doing this manually, and I wanted to share how the extraction pipeline actually works under the hood — and why "AI video summarizer" was the wrong way to think about the problem from the start. The first mistake I made was treating this as a summarization problem. Early prototypes fed the full YouTube transcript into an LLM and asked for "a summary of the tools used." The results looked plausible and were frequently wrong — the model would infer a database was PostgreSQL because the creator said "SQL," or assume a deployment target was Vercel because the video mentioned Next.js, even when the actual clip named Railway. Summarization models are optimized to sound coherent, not to be strictly grounded in the source. For a tool whose entire value proposition is "tell me exactly what was mentioned," a confidently wrong answer is worse than no answer. That's the core design constraint that shaped everything else: extraction, not interpretation. The pipeline itself is simpler than people expect. VidScope pulls the video's transcript (falling back to audio transcription when no captions exist), chunks it to stay inside context limits for long-form content — some tutorials and conference talks run past two hours — and runs extraction passes that are explicitly instructed to only output tool, framework, or library names that appear verbatim or near-verbatim in the transcript text. No inference from context, no "this is probably React because they mentioned JSX." If a name isn't in the transcript, it doesn't make it into the report. This is the "0% hallucination" constraint, and it's enforced at the prompt level with strict extraction rules, not just requested politely. The part that took the most iteration wasn't the extraction — it was verification. A raw list of tool names is only marginally more useful than pausing the video yourself. What makes the report actually save time is that every extracted name gets resolved to its official documentation or homepage link automatically, so a developer can go from "watched half a tutorial" to "have the actual docs open in five tabs" in under 30 seconds. Getting this resolution step accurate meant handling ambiguous names (there are at least four different open-source projects called "Flow"), filtering out generic terms that get mentioned but aren't tools (the model initially flagged "the cloud" and "database" as extracted entities before I tightened the extraction schema), and deduplicating variants of the same tool referenced multiple times under slightly different names across a single video. Long videos exposed a separate scaling problem: a two-hour conference talk produces a transcript that's tens of thousands of tokens, and naively extracting per-chunk without merging state across chunks produces duplicate or fragmented tool lists — the same framework mentioned in minute 5 and minute 90 shows up twice with slightly different context. The fix was maintaining an extraction state across chunks rather than treating each chunk independently, which also meant the final report generation step needed to be idempotent regardless of how many chunks a video's transcript happened to split into. This is the unglamorous part of building an "AI feature" that nobody talks about: the actual model call is a small fraction of the engineering effort compared to chunking, state management, and making outputs consistent at different video lengths. None of this matters if the tool doesn't fit into how developers actually work. The workflow I optimized for is: paste a YouTube link, wait 30 seconds, get a clean list of every tool and framework mentioned with verified links, export to Markdown or PDF, move on. No account required to try it, no video re-upload, no watching required at all. It's built for the moment when you find a promising tutorial at 11pm and want to know in thirty seconds whether it's using a stack you actually want to learn, before committing 45 minutes to watching it. If you've ever kept a messy notes file of "tools mentioned in videos I watched," or rewound a stream three times to catch a package name, that's exactly the problem VidScope solves. You can try it free on two videos, no credit card required, at vidscope.fr.