Prompting Techniques

From zero-shot to chain-of-thought, ReAct, and structured outputs.

Prompting a model to reason step-by-step before giving a final answer, which improves accuracy on multi-step problems.

e.g. "Let's think step by step..."

Providing a handful of example input/output pairs in the prompt to steer the model's behavior before asking it to do the task.

e.g. Three sample Q&A pairs followed by a new question.

A prompting pattern that interleaves reasoning ("Thought") with actions ("Act") like tool calls, then observes results before continuing.

e.g. Thought → search the web → Observation → Answer.

Constraining a model's response to a defined schema (like JSON) so downstream code can parse it reliably.

e.g. Forcing a response to match { "name": string, "age": number }.

A privileged instruction set at the start of a conversation that shapes a model's persona, rules, and constraints for the whole session.

e.g. "You are a concise, professional support agent."

Asking a model to perform a task from instructions alone, with no examples provided.

e.g. "Translate this sentence to French."