About LLM Prompting Techniques

Basic to advanced prompting techniques — chain-of-thought, few-shot, system prompts, patterns. Part of the DevTools Surf developer suite. Browse more tools in the Cheatsheets collection.

Use Cases

• Developers integrating LLM APIs and optimizing prompt quality
• Product managers writing system prompts for AI-powered features
• Content creators refining prompts for consistent tone and format
• Data scientists evaluating prompt strategies for classification tasks

Tips

• Check the anti-patterns section to avoid common prompting mistakes
• Use the structured output techniques for reliable JSON responses
• Review the chain-of-thought section for complex reasoning tasks

Fun Facts

• The term 'prompt engineering' barely existed before 2020; Google Trends shows near-zero search volume until GPT-3's release in June 2020.
• Chain-of-thought prompting, published by Wei et al. at Google Brain in January 2022, improved math reasoning accuracy by up to 50% on benchmarks.
• Few-shot prompting was demonstrated in the GPT-3 paper (Brown et al., 2020), showing that examples in the prompt could replace fine-tuning entirely.

FAQ

What's actually in here?

Prompt patterns that improve output: few-shot, chain-of-thought, role assignment, XML tagging, structured output. Plus anti-patterns that waste tokens.

Is this model-agnostic?

Mostly. Works across Claude, GPT, Gemini, Llama. Some techniques (Claude-specific XML tagging) are noted where relevant.

What about prompt injection?

Touched on — not a full security treatise. Focus is on getting good output, not defending against malicious inputs.

Does it cover RAG?

Brief. RAG is a system-level pattern (retrieval + prompting); this cheatsheet focuses on the prompting half.