Today Anthropic’s Zach Witten takes us on a deep dive into Anthropic’s cutting-edge AI models—Claude Haiku, Sonnet, and Opus—exploring their safety-first approach to generative AI and sharing essential tips for prompt engineering.

Topics Include:

• Introductions, about Anthropic
• 3 models: Haiku, Sonnet and Opus
• Scaling laws for hardware, data and compute
• Competing to be safest AI solutions, safety-first organization
• Leader in jailbreak resistance
• Interpretability features and breakthroughs for AI models
• Basics of prompt engineering
• Improving prompts with Claude
• Details matter – small changes to spelling, context will greatly improve results
• System prompt – role setting will improve results (i.e. “You are an expert mathematician…” for math query
• Be clear and direct – use XML tags where possible
• Encourage Claude to think step-by-step – answering fast comes with accuracy risk
• Use examples to provide additional clarity to Claude
• Bonus tips for image-based prompt engineering
• Q&A 1) Who wrote the meta-prompts in the cookbook?
• Q&A 2) Guidance for writing prompts for prompt generator
• Q&A 3) Best practices for tabular and structured data
• Q&A 4) Maintaining “tone” across hundreds/thousands of responses
• Q&A 5) Reverse engineering a prompt