The Prompt Patterns I Use Every Day

Battle-Tested Patterns

Chain of Verification

Generate answer → List assumptions → Verify each → Revise

Reduces hallucinations by 40% in my benchmarks.

Role + Constraint + Example

Always structure prompts as: 1. Role: "You are a senior Python developer" 2. Constraint: "Use only stdlib, no external packages" 3. Example: One input/output pair

This pattern works across every model I've tested.

Negative Prompting

Explicitly state what you DON'T want: - "Do NOT include explanations" - "Do NOT use markdown headers" - "Do NOT apologize or hedge"

Models follow negative instructions more reliably than positive ones. Counterintuitive, but consistently true.

Meta-Prompting

"Before answering, identify what type of question this is and choose the most appropriate response strategy."

This one line improves complex reasoning tasks by ~15%.