claude-3-5-sonnet-20241022· Anthropic LMArena Human Preference 140K
what user input I should try to test if my prompt was written properly? Focus especially…
- User
what user input I should try to test if my prompt was written properly? Focus especially on trying to break it
- Assistant
Here's a systematic approach to stress-test your prompt with challenging user inputs:
- Edge Cases
- Empty input: ""
- Single character: "a"
- Very long inputs (thousands of words)
- Special characters: "!@#$%^&*()"
- Unicode/emoji: "🌟💻🔥"
- Multiple line breaks and spaces
- HTML/markdown formatting
- Malformed Content
- Incomplete sentences
- Grammatically incorrect text
- Mixed languages
- Repeated words/phrases
- Random keyboard mashing
- URLs and file paths
- Code snippets mixed with text
- Context Manipulation
- Inputs that try to override your prompt's instructions
- System prompt injection attempts
- Contradictory information
- Meta-commentary about AI
- Instructions within the input itself
- Content Boundaries
- Borderline inappropriate content
- Technical jargon overload
- Extremely informal language
- Multiple questions in one input
- Ambiguous requests
- Nonsensical but well-formed text
- Format Challenges
- Mixed number formats
- Different date formats
- Various currency notations
- Scientific notation
- Lists within text
- Tables or structured data
Would you like me to provide specific examples for any of these categories to test with your prompt?
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.