YouTube transcripts → ChatGPT & Claude_
A working playbook for feeding YouTube transcripts into large language models — with prompt templates, chunking strategy, and a batch workflow for repurposing at scale.
1. Get a clean transcript
Start in the transcript extractor. Paste any public YouTube URL and export the transcript as plain text — strip timestamps if you plan to feed it straight into an LLM (timestamps eat tokens without adding meaning for most prompts).
2. Chunk for the context window
A one-hour video is typically 8–12k tokens. ChatGPT (GPT-4o) and Claude 3.5 Sonnet both handle that in one shot, but quality drops fast past ~15k tokens. Split long transcripts into 6–8k token chunks with 200-token overlap, then summarize each chunk and merge the summaries.
3. Prompt templates that work
You are an executive editor. Summarize the transcript below into 5 bullets, each ≤ 20 words. Preserve numbers, names, and dates verbatim. TRANSCRIPT: <paste chunk>
Rewrite the transcript as an 800-word blog post with an H1, three H2 sections, and a closing CTA. Keep the author's tone. Do not invent facts. TRANSCRIPT: <paste chunk>
Extract every non-obvious insight from the transcript as a JSON array of { insight, timestamp, quote }. Return only JSON.
TRANSCRIPT:
<paste chunk>4. Batch across many videos
For content repurposing at scale, use the batch pipeline to run the same prompt over up to 1,000 videos in a single job. This is how creators turn a full podcast back-catalog into a searchable knowledge base — or how researchers extract quotes from hundreds of interviews at once.
Pair this with the AI summarizer if you'd rather skip prompt engineering and get structured output straight away.