ai worlds fair AI LLM
Anthropic Prompt Doctor
- People bring prompts with Zack Witten
- Anthropic's workbench has eval built in
- Tips
- Use capitalization and proper grammar
- XML to separate parts of the prompt
- Claude is trained with XML data
- The most important thing is clearly separating sections of the prompt (other separators might work just as good)
- Helpful to avoid being too constraining (eg. "Every response must be..." vs. "Responses should be 3-4 sentences...")
- Be clear and reduce ambiguity
- Ordering of components?
- Info > Instructions
- Instructions are followed more closely to the end of the prompt
- Instructions after long documents
- Most reliable way to get Claude to output JSON is the assistant prefill functionality
- Putting words in Claude's mouth
- Utilize stop sequences!
- Claude seems to follow instructions better in the user message