There is the prompt you type, and there is the instruction layer sitting underneath it that you never see. Bani Chaudhuri walks through the second one. A system prompt is set once by whoever builds the agent, it arrives before anything the user says, and it frames how the model reads everything that follows.

The practical case is the one most people recognize. You use a chat tool for the same job every week and retype the same tone, audience, and length rules every time, until the conversation gets long enough that those early instructions stop carrying weight. Setting them as a system prompt instead means every conversation starts with the rules already loaded. The second half is a worked example, a content strategist prompt built up from a one-line role into role, tone, audience, format rules, and style rules, so you can see what the extra detail buys you.

What this video covers

• How a system prompt differs from the user prompt you type into a chat
• Why messages carry roles, and why a model follows system instructions more strictly than user ones
• Why repeating your context in every new chat stops working as the conversation grows
• A system prompt built out step by step, from a one-line role to a full set of rules
• Four habits that keep a system prompt working, review it, structure it under headings, test it against awkward questions, and start simple

Chapters

• 0:00 Why your AI keeps forgetting
• 1:21 How system prompts actually work
• 2:36 Turning a general chat tool into a specialized one
• 4:36 A system prompt template
• 6:09 Tips for building better agents