Prompt engineering has evolved significantly. The techniques that worked with GPT-3.5 are often suboptimal for GPT-5 and Claude 4. Here are 12 techniques that produce the best results with current frontier models.
1. Be Explicit About Reasoning Steps
Instead of “solve this problem,” say: “First, identify the key variables. Second, consider the constraints. Third, evaluate possible solutions. Fourth, select the best option and explain why.”
2. Use Structured Output Formats
Specify your output format precisely: “Return a JSON object with keys: summary (string), confidence (0-1), sources (array of strings).” GPT-5 follows these instructions with near-perfect reliability.
3. Provide Examples (Few-Shot) Sparingly
GPT-5 needs fewer examples than previous models. One good example is often sufficient. Three is the maximum before diminishing returns.
4. Use System Prompts Strategically
Put your role and constraints in the system prompt, not the user message. GPT-5 gives significantly more weight to system prompt instructions.
5. Ask the Model to Think Aloud
Adding “think step by step” or “show your reasoning” still works, but it’s more effective to say “before answering, explain your reasoning process in detail.”
6. Use Negative Instructions
Tell the model what NOT to do: “Do not use jargon. Do not exceed 200 words. Do not speculate beyond the provided context.”
7-12. Chain-of-Thought, Self-Consistency, Tree-of-Thought, ReAct, Refinement, and Meta-Prompting
These advanced techniques are detailed in the full article with code examples and benchmark comparisons. The key insight: with GPT-5, the gap between naive prompting and expert prompting has narrowed significantly, but expert prompting still produces 15-20% better results on complex tasks.