AI Terms, Explained Simply

The input you give an AI model — an instruction, question, or context — that guides its output. The quality of the prompt largely determines the quality of the response.

"Write a 200-word summary of this article for a beginner audience" is a prompt.

The practice of designing and refining prompts to get reliable, high-quality output from AI models. It combines task design, context, format control, and iterative testing.

Top-level instructions that define an AI's identity, capabilities, constraints, and rules for an entire conversation. It acts like a constitution governing every response.

A chatbot's system prompt might say: "Answer only from the provided knowledge base and never give legal advice."

The maximum amount of text (tokens) a model can consider at once, including the prompt, conversation history, and its response. Larger windows let you work with longer documents.

Claude's 200K-token context can hold a book-length document.

The basic unit of text a model processes. A token is roughly 3-4 characters or about 3/4 of a word. Models bill and limit input/output in tokens.

"Hello world" is about 2-3 tokens depending on the tokenizer.

A parameter controlling output randomness. Lower temperature (0-0.3) gives focused, deterministic answers; higher temperature (0.7-1+) gives creative, varied output.

Prompting a model to perform a task with no examples provided. The model relies entirely on its training. Works best for common, well-understood tasks.

"Classify this review as positive or negative:" with no examples.

Providing several input-output examples in the prompt to teach the model a task, format, or style before making the real request.

Showing three email→summary examples before asking for a fourth summary.

Providing exactly one input-output example to guide the model before the real request.

Prompting the model to reason step by step before answering. Improves accuracy on math, logic, and multi-step problems and makes reasoning auditable.

"Think step by step" before solving a word problem.

An advanced reasoning technique that explores multiple candidate reasoning paths, evaluates them, prunes weak branches, and deepens promising ones.

Reasoning + Acting. A prompting pattern that interleaves reasoning steps (Thought) with tool use (Action) and observation, powering AI agents.

Thought: I need the current price. Action: search("price of X"). Observation: result. Then answer.

An AI system given a goal, tools, and operating rules that can plan, execute actions, observe results, and iterate autonomously to achieve the goal.

The ability of a model to call external functions — search, calculators, APIs, code execution — and incorporate their results into its answer.

A confident, incorrect response generated by a model. Hallucinations occur when the model overgeneralizes patterns rather than recalling facts.

Restricting a model's answers to specific provided content (documents, data) rather than its general training knowledge, reducing hallucination.

NotebookLM answers only from your uploaded sources.

A pattern that retrieves relevant documents or data for a question, injects them into the prompt, and generates an answer grounded in them.

A numerical vector representation of text that captures semantic meaning. Embeddings power semantic search and similarity matching between texts.

The process by which a model selects the next token based on predicted probabilities. Sampling parameters like temperature and top-p shape creativity and determinism.

A sampling parameter that limits choices to the smallest set of tokens whose combined probability exceeds P. Lower values make output more focused.

A model's ability to learn a task from examples given in the prompt without updating its weights. This is what makes few-shot prompting work.

Models that can process multiple input types — text, images, audio, and video — often within a single prompt.

GPT-4o and Gemini accept text plus images together.

Further training a pretrained model on a specific dataset to specialize its behavior. Distinct from prompting, which does not change weights.

Rules and constraints that keep an AI system safe and aligned: prohibited topics, output limits, refusal conditions, and safety filters.

A technique where hidden instructions in user content try to override the system prompt or cause unintended behavior. A key security concern for AI applications.

A neural network trained on vast text data to predict and generate language. LLMs are the engines behind modern chatbots and AI assistants.

GPT-4o, Claude, and Llama are LLMs.

A model whose trained parameters are publicly released, allowing self-hosting, fine-tuning, and customization. Examples: Llama, DeepSeek, Flux, Stable Diffusion.

Search that matches meaning rather than exact keywords, using embeddings to find conceptually related results even when wording differs.

A condensed reference of proven prompt formulas, structures, and patterns organized for quick use. AiClubb publishes free printable cheat sheets.

A reusable prompt skeleton with variables (placeholders) you fill in per use. Templates standardize quality across many similar requests.

A brand-voice template with fields for topic, audience, and tone.

A sequence of AI steps that together accomplish a goal, often using prompt chaining or agents. Workflows make complex tasks repeatable and reliable.