A prompting technique that has AI write out its reasoning step-by-step before giving a final answer. The core approach: add "think step by step" to prompts or provide reasoning step examples. Research shows this seemingly simple technique can dramatically improve large language models' accuracy on math, logical reasoning, and multi-step problems — sometimes from 20% to over 80%.

Providing a few "input→output" demonstration examples in the prompt, letting AI learn your expected format, style, or task pattern, then applying it to new inputs. More than Zero-Shot (stating requirements directly); much lighter than <a href="/en/glossary/core-concepts/fine-tuning/">Fine-Tuning</a> (retraining the model). When tasks have specific non-standard format requirements or need highly consistent output style, Few-Shot far outperforms relying on text description alone.

Prompt Compression is the practice of reducing the length of what you send to an AI model without losing critical information. The goal is to prevent long or multi-turn conversations from overflowing the model's <a href="/en/glossary/core-concepts/context-window/">Context Window</a> or generating unnecessary cost and latency. Common techniques include summarizing earlier exchanges into key points, keeping only the most recent turns in full, removing resolved side threads, and replacing long documents with distilled summaries.

The systematic design and optimization of instructions given to AI to produce more accurate, need-aligned outputs. The core skill that makes the same Claude model produce completely different quality results depending on how instructions are crafted.

An attack where malicious instructions are embedded in external content (web pages, documents, user inputs) to override or bypass an AI system's <a href="/en/glossary/prompt-techniques/system-prompt/">System Prompt</a>, causing the AI to execute unauthorized actions.

A "prompt" is the question or instruction you input in the conversation; a "<a href="/en/glossary/prompt-techniques/system-prompt/">System Prompt</a>" is the "background rules" set before the conversation starts, determining Claude's role, behavioral guidelines, and constraints in this conversation. Analogy: a prompt is you telling Claude "do this thing"; a <a href="/en/glossary/prompt-techniques/system-prompt/">System Prompt</a> is you telling it before the conversation "who you are and how you should behave." For regular claude.ai users, this difference is most visible in <a href="/en/glossary/claude-tools/claude-projects/">Claude Projects</a>' Instructions settings.

Assigning Claude a specific role or expert identity at the start of a prompt, causing it to respond using that role's knowledge framework, tone, and reasoning style. One of the most fundamental and effective prompting techniques.

A hidden instruction layer sent to Claude before conversation begins, used to define its role, set rules, and constrain behavior. All messages in the conversation operate within the <a href="/en/glossary/workspace-basics/system-prompt/" target="_blank">System Prompt</a>'s framework, and it is invisible to end users.

Giving AI a task instruction and expecting it to complete it without providing any examples — "zero demonstrations, just shoot." The opposite of <a href="/en/glossary/prompt-techniques/few-shot-prompting/">Few-Shot Prompting</a> (providing examples for AI to learn format from). Most people's everyday Claude usage is actually Zero-Shot: just saying "write me an email" without providing examples first.