Prompting 101

@hao CC BY 4.0

Video: https://www.youtube.com/watch?v=ysPbXH0LpIE

  • [00:00] Introduction to the session by Hannah and Christian from Anthropic's applied AI team.

  • [00:20] The session focuses on prompting best practices using a modified real-world scenario.

  • [02:10] Scenario description: Swedish insurance company analyzing car insurance claims.

  • [02:30] Information used: car accident report form and a human-drawn sketch.

  • [03:50] Use of Anthropic's Claude model with specific settings.

  • [04:05] Settings include temperature zero and a large max token budget.

  • [05:40] Best practices for prompt structure: setting task description up front.

  • [06:00] Include content, detailed instructions, examples, and repeat important information.

  • [08:30] Importance of task context and tone context in prompt engineering.

  • [09:00] Ensuring Claude stays factual and confident without guessing.

  • [11:20] Version 2 of the prompt has more detailed instructions.

  • [11:50] Claude identifies the scenario as a car accident and reflects vehicle involvement.

  • [13:40] Background information about the Swedish car accident form.

  • [14:00] Details like form structure and common markings explained.

  • [15:50] Use of delimiters like XML tags and markdown to organize information.

  • [16:10] XML tags help Claude reference information during task execution.

  • [20:10] Importance of examples or few-shot learning in steering Claude.

  • [20:30] Examples can include data and descriptions for better model understanding.

  • [22:00] Conversation history as a tool to enrich Claude's context.

  • [22:20] Useful for user-facing applications with ongoing interactions.

  • [24:30] Final task reminder and guidelines to prevent hallucinations.

  • [24:50] Claude is instructed to answer only when confident, with references to data.

  • [28:30] Output formatting ensures data is stored correctly for applications.

  • [28:50] Use of XML tags for final verdicts to integrate with systems.

  • [31:20] Prefilled responses to shape Claude's output format.

  • [31:40] Ensures structured JSON output for serializable data handling.

  • [33:10] Claude 3.7 and 4 as hybrid reasoning models with extended thinking.

  • [33:30] Extended thinking allows analysis of Claude's reasoning process.