Why pasting the link fails
ChatGPT cannot watch a video. Depending on the model, the mode and whether browsing is on, a YouTube link either gets fetched as a normal web page, which mostly yields the title and description, or it does not get fetched at all.
The dangerous part is what happens next. Instead of saying it could not read the video, the model may produce a summary built from the title, the channel and its own training data. It reads perfectly. It is also, on any specific claim, unverifiable. People have quoted numbers from summaries of videos the model never opened.
Feeding the transcript removes the guesswork. The words in the chat are the words in the video, so the output can be checked line by line.
Getting the text out, fast
Manually, it goes: expand the description, click Show transcript, open the three-dot menu, toggle timestamps off, click before the first word, scroll to the end, shift-click, copy. Seven steps, and step six is a fight with auto-scroll on anything over an hour.
The short path is a copy button. Our transcript extractor puts one in the side panel: click the icon, click Copy, paste in the chat. Timestamps come along or stay behind depending on a toggle.
Drop the timestamps before pasting into a chat. They roughly double the token count, they add nothing the model can use, and they push a long video over the limit faster. Keep them only if you plan to ask for time-coded references back, which is a legitimate but much rarer case.
Prompts that produce something useful
Instruction first, then the text, with an obvious separator. Models follow the top of a long message better than the bottom.
Summary with structure: "Below is the transcript of a talk. Give me: the main argument in two sentences, then the three claims it rests on, then anything the speaker admits they are unsure about. Quote exact wording for the claims. Transcript follows after the line of dashes."
Two things make this beat "summarize this". Asking for exact quotes keeps the model anchored to the text. Asking what the speaker was unsure about surfaces the hedges a normal summary flattens into confidence.
Other requests worth having ready: pull every tool, product or person mentioned into a list; rewrite the argument as an outline for a blog post; extract only the section about a particular topic and ignore the rest; list the claims that would need a source before publishing.
For study notes, ask for questions rather than statements. "Turn this lecture into twelve questions with answers, ordered as the lecturer presented them" produces something you can revise from.
When the transcript is too long
A two-hour podcast runs roughly 20,000 words, which is well inside modern context windows but will still hit limits on smaller models or long chat histories. Options, in the order worth trying:
- Strip timestamps. Often enough on its own, and free.
- Start a fresh chat. A long history eats the same budget as the transcript.
- Attach it as a file. Paste the .txt export rather than the message body where the interface supports it.
- Cut to the part you need. Search the transcript for your term, take the surrounding section, paste that. Output usually improves too, since the model stops averaging over ninety minutes of unrelated material.
Chunk-and-merge, where you summarize part one, summarize part two, then summarize the summaries, is the last resort. It reliably loses arguments that run across the whole video, which are usually the arguments worth having.
When to skip ChatGPT entirely
If the only goal is "tell me if this video is worth watching", the round trip through a chat window is overhead. A summarizer built into the video page reads the transcript and answers that in one click. Ours streams from Gemini 2.5, and the summarizer page covers what it does and does not do well.
Keep ChatGPT for the jobs where you have a specific question, an unusual output format, or a follow-up conversation in mind. Triage does not need a chat.