How to summarize a YouTube video with AI.

The reliable way to summarize a YouTube video with AI is to start from the transcript that's already on the watch page and ask the model for the output depth you actually need — bullets for triage, timestamps for review, full notes for study. Anya uses this to scan 90-minute industry panels in 60 seconds. Cleo turns recorded lectures into exportable study notes. This guide walks through the working method end-to-end, plus what to do when the video has no captions, when the transcript is too long for one request, and when the AI provider rate-limits your fifth summary of the day.

Setup steps

1. 1

   Open the YouTube video in Chrome and wait for it to load

   Navigate to the video on YouTube. Let the page finish loading — captions and metadata sometimes lag the video itself. On slow connections, give it 5-10 seconds before triggering the summarizer.

2. 2

   Confirm captions or a transcript exist

   Click the CC (captions) button or the gear icon to see the subtitle/CC menu. If the menu shows no caption tracks (not even auto-generated), there's no transcript to summarize. Wait a few minutes if the video is brand new — YouTube's auto-captioning runs asynchronously.

3. 3

   Pick the output style before clicking summarize

   Decide what you want: a concise bullet summary for triage, a full notes version for study, a timestamped Moments artifact for review, or a social-post draft for sharing. Reduz exposes these as output styles in extension settings so you don't have to write prompts each time.

4. 4

   Click Reduz on the active YouTube tab

   Open the Reduz extension from the toolbar (or click the injected Summarize button on the YouTube watch page). The extension reads the transcript directly from the page — no copy-paste, no separate web tool.

5. 5

   For long videos, use selected text instead

   For 3+ hour podcasts and panels, the full transcript may exceed the AI provider's the model can fit. Open YouTube's transcript view, select the section that matters, right-click "Summarize text with Reduz." Only that section becomes the source.

6. 6

   Switch providers if the AI side fails

   If transcript extraction worked but the summary itself failed, the issue is the AI provider (capacity, rate limit, quota). Switch Reduz to a different provider — Claude, Gemini, DeepSeek, or Grok — and retry. The same transcript runs through a different model.

7. 7

   Save and export the useful output

   Every summary saves to local storage history on your device by default. Export to Markdown, PDF, DOCX, JSON, or a ZIP bundle when you need to share it or move it into another note-taking tool.

Transcript quality matters more than model choice

AI cannot recover detail that's missing from the source text. If YouTube's auto-captions misheard "Anthropic" as "ant-stropic," that error propagates into your summary. Quality of input matters: human-uploaded subtitles are usually clean; auto-captions on clear English audio are usually good; auto-captions on heavy accents, music tracks, or technical jargon are often noisy. Before blaming the AI for a bad summary, check the source transcript. You can view it in YouTube's native transcript panel (three-dots menu under the video title).

Output styles vs writing custom prompts

The same transcript can produce bullets for triage, full notes for study, timestamped Moments for review, a social post for sharing, or an action-items list for meeting recordings. Writing a new prompt every time is slow and inconsistent. Reduz exposes output styles in settings — pick once, generate repeatedly. For unusual cases (a very specific output structure for a research project), Reduz also supports custom prompts with placeholders like `{title}`, `{url}`, and `{content}`.

When the full transcript is too long

A 3-hour podcast can produce a 30,000-token transcript — enough to exceed smaller models' the model can fits. Three workarounds: (1) select-text mode on the section that matters; (2) switch to a larger-context model (Claude Sonnet 4.6 or GPT-5.5 handle long transcripts comfortably); (3) generate chunked summaries by selecting different time ranges and summarizing each. For weekly podcasts you watch in their entirety, larger-context models are usually the right default.

Privacy on YouTube transcripts

YouTube transcripts are visible to anyone watching the video — they're not sensitive data themselves. The privacy question is about prompts and outputs: where does the transcript go for AI generation? In Reduz bring your own AI key, the transcript goes direct from your browser from your tab to your selected AI provider using your own key. No Reduz server, no vendor relay. In Hosted Free, the transcript goes through the Reduz relay (with an installation identifier for credit counting only) to a hosted model. For most YouTube summarization, either path is acceptable.

When AI summaries are not enough

For research, citations, and quotes, the AI summary is a starting point — verify exact claims, numbers, and direct quotes in the original video. AI models occasionally hallucinate specific numbers or attribute statements to the wrong speaker. For time-stamped reference in a literature review or a citation, the AI summary points you to the moment; the actual quote should come from your own listening or YouTube's transcript. Treat AI summaries as triage and prep, not as the source of record.

Practical checklist

• Confirm the video has usable captions (CC button shows a track) before attempting summarization.
• Pick an output style — bullets, full notes, Moments, social post — before clicking summarize.
• For 3+ hour videos, use selected-text mode on the section that matters, not the full transcript.
• Switch to a larger-context model (Claude Sonnet 4.6 or GPT-5.5) if smaller models reject the transcript size.
• Use bring your own AI key when you don't want even the transcript text touching a vendor relay.
• Verify direct quotes and specific numbers against the original video before citing.
• Save useful summaries to local history and export to Markdown for revisit weeks later.

Frequently asked questions