DEV Community

Cover image for How to summarize 2-hour videos and 50-page articles without leaving the browser tab
glensum
glensum

Posted on

How to summarize 2-hour videos and 50-page articles without leaving the browser tab

A 40-minute conference talk. A 90-minute podcast. A 12-minute tutorial that could have been a blog post. Most of what's valuable in a video is its content — and content can be read.

I've been building GlenSum, a browser extension that summarizes long videos and articles directly in the tab. This post covers the three working approaches to media summarization in a browser, the privacy trade-off most people don't notice they're making, and what "long content" actually means for an LLM.

Three ways to summarize a YouTube video (slowest to fastest)

1. The transcript, by hand

Every video with captions has a transcript: expand the description → Show transcript → select all → paste into any AI chat.

  • Cost: free
  • Time: 3–5 minutes of fiddly copying, and raw transcripts are messy — no punctuation, no speaker labels, timestamps everywhere
  • Best for: one-off use when you can't install anything

2. Paste the URL into an AI chat

Some assistants accept a YouTube URL directly and fetch the transcript themselves. Paste the link, ask for key points.

  • Cost: often limited by your plan's usage quota
  • Caveat: availability varies by model and region — and now the video passes through yet another service
  • Best for: occasional videos when you already pay for an assistant

3. A browser extension, one click

Extensions built for this read the transcript from the page you already have open and generate a structured summary in a side panel. The workflow becomes: open the video → click the extension → read the summary. No copying, no tab-switching.

The extras that matter: adjustable summary length, follow-up questions about the content ("what did they say about X?"), bilingual output when the talk isn't in your first language, and Markdown/PDF export so the summary lands in your notes.

The privacy trade-off most people miss

Here's the part that motivated building this in the first place. When you paste a URL or transcript into a hosted summarizer, your reading and watching history becomes someone's dataset. Article by article, video by video, a profile of what you consume accumulates on a server you don't control. That's a real cost that never shows up on the pricing page.

The alternative is BYOK — bring your own key:

  1. You paste an API key from your own AI provider (OpenAI, Google, a local gateway, anything OpenAI-compatible) into the extension once.
  2. The extension reads the page content in your browser and sends it directly to the provider you chose.
  3. There is no middleman server. No account. No logs on my side, because there is no "my side" — the extension ships as static code, and the network calls go from your browser to your provider.

The trade-off is honest: you manage your own key, and free-tier users can still use built-in free models without one. But nobody in the middle ever sees what you read.

What "long" actually means for an LLM

Summarizing a 2-hour talk isn't just "paste transcript into prompt". A two-hour transcript is roughly 25,000–40,000 words — technically it may fit a modern context window, but quality degrades: the model skims, drops the middle, and blends sections together.

The approach that works is map-reduce over chunks:

  • Map: split the transcript (or article) into overlapping chunks at natural boundaries — chapters, speaker turns, section headings — and summarize each chunk separately.
  • Reduce: merge the chunk summaries into a final structured summary, keeping section-level detail instead of a mushy average.

This is also why page-native tools beat copy-paste: the extension already knows the page structure, so it can split at real boundaries instead of arbitrary token counts.

When a summary is the wrong tool

Summaries are ideal for inverted-pyramid content: talks, podcasts, news analysis, tutorials — where information density is low and predictable. They're poor substitutes for demonstration content: if someone is showing you how to do a three-point turn, no summary captures the steering.

A good rule: summarize first to decide whether to watch — not to avoid watching things worth watching.

Try it

GlenSum is free to install and works out of the box (built-in free models, or your own key for unlimited use):

Questions about the architecture — especially around BYOK flows or chunking strategies for long transcripts — happy to answer in the comments.

Top comments (0)