prompt-engineering

Start with a simple direct instruction, then add constraints, reasoning steps, and 2-5 input-output examples only as needed. Test on diverse inputs including edge cases, and iterate using A/B comparisons to measure accuracy and consistency.

Few-shot prompting teaches the model by showing 2-5 input-output example pairs instead of explaining rules. Use it when you need consistent formatting, specific reasoning patterns, or edge-case handling, balancing example count against token cost.

Chain-of-thought prompting requests step-by-step reasoning before the final answer, either with a phrase like "Let's think step by step" or example reasoning traces. It improves accuracy on multi-step analytical and mathematical tasks by 30-50%.

System prompts should hold stable instructions: the model's role, expertise level, output format, and safety guidelines. User messages should carry variable per-turn content, which keeps behavior consistent and frees user-message tokens.

Inconsistency usually comes from vague instructions, missing output format constraints, or examples that don't match the target task. Add explicit constraints, include representative examples, and test on boundary inputs to stabilize outputs.

Common pitfalls include over-engineering prompts before trying simple versions, exceeding token limits with excessive examples, leaving ambiguous instructions open to multiple interpretations, and never testing edge cases.