HN Explorer

Show HN: I built a tool to clear my YouTube's "Watch Later" Video Graveyard (recapio.com)

2 points by nikhonit ↗ HN
I built this because I consume a lot of technical lectures and long-form podcasts on YouTube, but I found myself wasting hours scrubbing through videos just to find specific citations or concepts.

Recapio is a tool that extracts the transcript and generates structured summaries for videos (and web articles). It’s not trying to replace watching content, but rather to act as a 'Ctrl+F' for video context.

One technical challenge I faced: Dealing with auto-generated YouTube captions vs. forced captions was messy. I had to build a parser that normalizes the timestamps so that when you click a summary point, it actually seeks to the correct frame, even if the caption timing is drifting.

It has a free tier that should cover most casual usage. I’d love your feedback on the extraction quality

3 comments

[ 2.7 ms ] story [ 23.5 ms ] thread
nikhonit
I built this because my YouTube 'Watch Later' playlist hit 500+ videos and became a source of anxiety rather than a queue. I wanted to know if a 2-hour podcast contained the specific citation I needed without watching the whole thing.

Under the hood: Recapio grabs the transcript (prioritizing manual captions over auto-generated ones) and uses an LLM to generate structured summaries with timestamped citations.

The Engineering Challenge: The biggest headache was 'hallucination drift'—where the AI summary claims a topic starts at 10:00, but it actually starts at 10:45. I solved this by implementing a chunking strategy that overlaps context windows, forcing the model to verify timestamps against the raw text segments before outputting the link.

It’s a work in progress. I'm curious if anyone has better strategies for handling the lack of punctuation in auto-generated YouTube captions

nikhonit
A quick note for the developers/hackers here:

While building this, I realized most YouTube transcript APIs were either overpriced or lacked good integration for LLM workflows.

So I spun out the backend as a standalone API: transcriptapi.com

The cool part is I added native MCP (Model Context Protocol) support. If you use Claude Desktop or similar agents, you can drop this in as a tool to fetch full video context directly into your chat window without copy-pasting.