Why long videos break the usual methods
Three different ceilings, and they fail quietly rather than loudly:
- Message limits. Pasting a two-hour transcript into a chat window does not fit. You split it into parts, and the model loses the connection between something said at 00:12 and the thing that contradicts it at 01:47.
- Truncation. Many tools cap what they send to the model. The cap is rarely disclosed and the output never mentions it.
- Sampling. Some take the beginning, a slice of the middle and the end. For a structured lecture that is often fine. For an interview where the real answer arrives at minute 80, it is useless.
None of these produce an error. They produce a confident summary of part of a video, presented as a summary of the video.
How to tell you got a partial summary
Two checks, both about ten seconds.
Look at the last timestamp cited. If the newest moment referenced sits at 22 minutes into a 95-minute video, you have a summary of the first quarter. This only works if your tool cites timestamps at all — if it does not, you have no way to run this check, which is itself worth knowing.
Ask about the end. “What did the speaker conclude in the last fifteen minutes?” A tool working from the whole transcript answers specifically. A tool working from the first slice either hedges or invents something plausible.
A tool that stops and says it could not read the whole video is doing something the others do not: admitting a limit. Fluent output is not evidence of complete input, and for a long video that gap is where most wrong summaries come from.
Prompts that suit a long video
Asking for “a summary” of three hours gets you something too flat to use. These work better:
- Build a chapter map. “Break this into sections with timestamps, one line each, so I can decide what to actually watch.”
- Track one thread. “Follow what the speaker says about pricing across the whole video, in order, with timestamps.”
- Find the turn. “Where does the argument change direction, and what prompted it?”
- Catch the contradiction. “Does the speaker say anything late in the video that conflicts with what they said early on?” This is the question a partial summary can never answer.
- Skip to your part. “I only care about the section on hiring. Where is it and what does it say?”
The honest limit
The video still needs captions on YouTube — a three-hour upload with none cannot be read by this or anything else, and the plugin will tell you so rather than produce something. Very long videos also take longer to read before the first answer comes back; there is no way around that, the text has to be retrieved.
Related: when a tool says there is no transcript but the video has one, and what ChatGPT can and cannot do with a video.