A lightweight application that uses AI to summarize provided videos.
Video Brief
Video Brief turns long YouTube videos into concise, timestamped summaries that are easier to scan, reference, and revisit.
Paste in a public video with captions and it breaks the content into key sections, highlights the main ideas, and links each section back to the relevant point in the source video.
Built as a small, focused tool for getting the useful parts of a video without having to scrub through the entire thing.
1. Get The Transciript
Using yt-dlp, I am able to query YouTube's API and pull the transcript of any public video via the URL. Transcripts are cached permenantly to reduce repeat queries to YouTube.
2. Summarize
The transcript is then sent to an LLM for summarization. Since this is a simple, repeatable task, I chose to use qwen3.5:9b and provided it with a response template to follow.
3. Deliver
The LLM responds with a markdown formatted summary, then the site renders the summary and reformats any timestamps as links to the specified video segment. Summaries can be downloaded, shared, and revisited for 48hrs.
4. Data Retention
To strike a balance between avoiding rate-limits and storing as little data as possible, I opted to preserve pulled transcripts indefinitely and only retain generated summaries for 48hrs. This reduces the chances of youtube rate-limiting my connection and allows videos that have already been pulled to skip straight to summarization. Allowing summaries to be deleted also allows them to be regenerated with the latest prompt and model selections behind the site, which means that summaries of cached transcripts can improve over time.