I’m building a Chrome extension that embeds a chat panel next to any YouTube video. This chat allows viewers to ask questions like “Summarize this video and give me the important timestamps,” and the model responds with context-aware answers.

For each video, I collect the transcript, description, and metadata (e.g., likes, title, duration), and feed all this information as a system message to ChatGPT. I also include another system message with formatting and behavioral rules. These rules can be quite extensive:

1. What you are and why you're doing this
2. Behaviour rules (responses should be X characters long, do not talk about things that are not in the video, etc)
3. Formatting rules (how to do bold, italics, lists, etc)
4. Common usecases and desired results

However, for longer videos (1+ hour), the transcript can be extremely large, and the combination of detailed context and numerous rules sometimes causes the model to produce confused or suboptimal responses.

Given that speed is crucial (I want to avoid multiple prompt iterations per message), what strategies or best practices can I use to optimize my prompts and ensure consistent, high-quality responses from the model?

Any advice or pointers would be greatly appreciated!

PS I'm using gpt-4o-mini (for the speed and good quality) with 0.3 temp.

I’m building a Chrome extension that embeds a chat panel next to any YouTube video. This chat allows viewers to ask questions like “Summarize this video and give me the important timestamps,” and the model responds with context-aware answers.

For each video, I collect the transcript, description, and metadata (e.g., likes, title, duration), and feed all this information as a system message to ChatGPT. I also include another system message with formatting and behavioral rules. These rules can be quite extensive:

1. What you are and why you're doing this
2. Behaviour rules (responses should be X characters long, do not talk about things that are not in the video, etc)
3. Formatting rules (how to do bold, italics, lists, etc)
4. Common usecases and desired results

However, for longer videos (1+ hour), the transcript can be extremely large, and the combination of detailed context and numerous rules sometimes causes the model to produce confused or suboptimal responses.

Given that speed is crucial (I want to avoid multiple prompt iterations per message), what strategies or best practices can I use to optimize my prompts and ensure consistent, high-quality responses from the model?

Any advice or pointers would be greatly appreciated!

PS I'm using gpt-4o-mini (for the speed and good quality) with 0.3 temp.

• openai-api
• gpt-4

1 Answer 1

I would start by looking at Retrieval Augmented Generation to include only the relevant parts of the video for a query instead of sending the transcript fully