System Prompt Tester

A system prompt is an instruction given to an LLM before the user conversation begins. It sets the AI's persona, tone, constraints, and capabilities. A well-crafted system prompt can dramatically improve response quality — turning a generic assistant into a domain expert with specific output formats, safety rules, and consistent behavior.

Best practices: (1) Define the role/persona clearly upfront. (2) Specify the output format explicitly (JSON, markdown, bullet points). (3) List constraints and what NOT to do. (4) Provide examples if needed. (5) Keep it concise — longer isn't always better. (6) Use XML tags for structured sections in Claude. (7) Test with edge cases, not just happy paths.

Few-shot prompting means including examples of input-output pairs in the system prompt to guide the model's behavior. Instead of just describing what you want, you show it. Example: 'User: analyze sentiment of: I love this. Assistant: Positive (confidence: 0.95)'. This dramatically improves consistency for structured tasks.

Test with: (1) Typical user inputs for your use case. (2) Edge cases and ambiguous requests. (3) Adversarial inputs that try to bypass constraints. (4) Different phrasings of the same request. (5) Very short and very long inputs. Compare outputs across test cases to find where the prompt breaks down and iterate.

Common mistakes: (1) Too vague — 'Be helpful' gives no real guidance. (2) Too restrictive — blocking legitimate use cases. (3) No output format specification — getting inconsistent structures. (4) No error handling instructions — model doesn't know what to do with invalid input. (5) Using examples that are too similar to each other, missing edge cases.