You instruct the model to reason step-by-step before giving a final answer. Instead of jumping to a conclusion, the model walks through its logic — and produces dramatically more accurate results for complex tasks. CoT forces the model into "System 2 thinking" — deliberate, analytical reasoning rather than pattern-matching.

You provide 2–5 examples of the desired input-output pattern before asking the model to complete a new one. The model learns the pattern and applies it. Few-shot is the most reliable way to enforce consistent output formatting.

ReAct interleaves reasoning traces with actions. The model thinks about what it needs to do, performs an action, observes the result, then reasons about the next step. It's the foundation of AI agent behavior — the architecture behind Claude Code's extended thinking and GPT-5's tool use.

You ask the same question multiple times, each with a different reasoning path, and take the most common answer. This filters out lucky guesses and random hallucinations by requiring convergence across approaches.

Instead of a single chain of reasoning, the model explores multiple thought branches simultaneously, evaluates each, and pursues the most promising ones — the AI equivalent of brainstorming on a whiteboard.

You provide a set of principles the model must follow when generating its response, and it self-critiques its output against them before presenting it. Anthropic's Claude models are natively trained with Constitutional AI.

You structure your prompt using XML-style tags to clearly delineate instructions, context, examples, and output format — Anthropic's officially recommended format. Claude parses XML-structured prompts with near-perfect accuracy.

You assign the model a specific role, expertise level, communication style, and perspective. This activates the model's knowledge in that domain and shapes its output style.

Instead of one massive prompt, you break the task into a sequence of smaller prompts, each building on the previous output. Chaining reduces errors, lets you inspect intermediate outputs, and allows course-correction between steps.

You explicitly list what the model should avoid — words, patterns, tones, structures. This is often more effective than only describing what you want, since models trained to be comprehensive sometimes over-explain or default to clichés.

You treat the AI as a collaborator in an editing process. Start with a draft, then refine through specific feedback. The best AI outputs rarely come from a single prompt — they come from a conversation.