AI Agent Use Cases
Views —

AI Agent Use Cases

Added today

10All 241

1–20 of 241

Work & ResearchWalkthrough

A Bot That Summarizes Watched YouTube Videos and Finds Specific Scenes

Send a link through Telegram and it summarizes the video, stores the transcript in one-minute chunks, and lets Claude find links to relevant scenes.

n8n

How it works

  1. The creator was using an n8n bot where sending a YouTube link with the /yt command in Telegram prompts GPT to summarize it, with Gemini taking over if GPT fails.
  2. Because past summaries were hard to find, they split transcripts into roughly one-minute chunks with timestamps, attached the video title, and stored them in a Qdrant database on the same server.
  3. They turned a second n8n workflow into an MCP server with a single search tool and connected it to Claude, which answers questions with links that jump to the relevant second in the video.

Results

The creator fixed an issue where transcripts were split twice and half the stored entries contained only the title by increasing the split size to 2,000 characters. They also changed it to delete and reinsert the same video, and mainly use GPT because the free Gemini service can take two to three minutes even for a 15-second video.

Notes

They said both workflows and setup instructions were posted on GitHub with credentials removed. No search-accuracy metric was provided.

Source

Reddit Original post ↗

Example prompt · n8n

Build an n8n bot that summarizes YouTube videos I have watched and, when I ask later, finds a link to the relevant scene.

Inputs: messenger for sending links [for example, a Telegram bot], summarization model [primary model and fallback model], database for storage [for example, Qdrant on my server], and place for asking questions [for example, a custom Claude connection]. First, verify whether transcripts can be retrieved with timestamps and whether the database can be used without exposing it publicly.

Implementation steps: 1) When a link arrives, retrieve the transcript, summarize it, and reply through the messenger. 2) Split the same transcript into roughly one-minute chunks and store each with the video title and start time. 3) Set the storage-stage split size larger than each chunk so chunks are not split again. 4) If the same video is submitted again, delete its existing chunks before inserting it. 5) Create a workflow with a single search tool that returns matching chunks and timestamped links for a question.

Deliverables: two workflows and setup instructions. First, add two videos, then verify that none of the stored chunks contains only a title and that adding the same video twice does not increase the chunk count.