AI Prompt Engineering Glossary

An AI system that can take actions autonomously to accomplish tasks. Agents can use tools, make decisions, and execute multi-step workflows without human intervention at each step.

A workflow where an AI model operates with increased autonomy, making decisions and taking actions to achieve a goal rather than just responding to single prompts.

The process of ensuring AI systems behave in accordance with human values and intentions. This includes safety measures, ethical guidelines, and preventing harmful outputs.

An AI safety company and the creator of Claude. Founded by former OpenAI researchers, Anthropic focuses on developing safe and beneficial AI systems.

A set of protocols that allows different software applications to communicate. AI APIs let developers integrate language models into their applications programmatically.

Processing multiple prompts or requests together rather than one at a time. Often more cost-effective for large-scale operations.

Systematic patterns in AI outputs that reflect prejudices or imbalances in training data. Can manifest as unfair treatment of certain groups or overrepresentation of certain viewpoints.

A prompting technique that encourages the model to break down complex reasoning into intermediate steps. This often improves accuracy on math, logic, and multi-step problems.

An AI assistant created by Anthropic. Known for being helpful, harmless, and honest. Available in different versions including Claude Instant and Claude Opus.

The text generated by an AI model in response to a prompt. Also called the "response" or "output."

An alignment technique developed by Anthropic where AI models are trained to follow a set of principles (a "constitution") that guide their behavior.

The maximum amount of text (measured in tokens) that a model can process in a single interaction. This includes both your prompt and the model response.

A structured prompt framework: Context, Objective, Style, Tone, Audience, Response format. Helps create comprehensive, effective prompts.

Producing the same output for the same input every time. AI models are non-deterministic by default but can be made more deterministic by setting temperature to 0.

A type of generative AI model that creates content by gradually removing noise from random data. Commonly used for image generation (like Stable Diffusion, DALL-E).

A numerical representation of text (or other data) as a vector of numbers. Embeddings capture semantic meaning, allowing computers to understand relationships between concepts.

Capabilities that arise in large models that were not explicitly programmed or anticipated. As models scale, they sometimes develop unexpected abilities.

A prompting technique where you provide several examples of the desired input-output pattern before your actual request. Helps the model understand the format and style you want.

The process of further training a pre-trained model on a specific dataset to specialize it for particular tasks or domains. More resource-intensive than prompt engineering.

A large AI model trained on broad data that can be adapted to many different tasks. GPT-4, Claude, and Llama are examples of foundation models.

AI systems that can create new content such as text, images, audio, or code. Large language models are a type of generative AI focused on text.

A family of large language models developed by OpenAI. GPT stands for the architecture (Generative Pre-trained Transformer) but is commonly used to refer to the models themselves.

Connecting AI responses to factual sources or real-world data. Grounding helps reduce hallucinations by anchoring outputs in verifiable information.

Safety mechanisms and constraints built into AI systems to prevent harmful, inappropriate, or off-topic outputs. Can be implemented through training, prompting, or output filtering.

When an AI model generates plausible-sounding but factually incorrect or fabricated information. A significant challenge with current language models.

The ability of language models to learn from examples provided within the prompt, without any weight updates or fine-tuning. The basis for few-shot prompting.

The process of running a trained model to generate outputs. When you send a prompt to an AI and get a response, that is inference.

A type of fine-tuning where models are trained to follow instructions more accurately. This makes models better at understanding and executing user requests.

A feature in some AI APIs that ensures the model output is valid JSON. Useful when you need structured, parseable responses for programmatic use.

Attempts to bypass an AI model safety measures and guardrails through creative prompting. AI providers continuously work to prevent jailbreaking attempts.

The date after which a model has no training data. Events and information after this date are unknown to the model unless provided in the prompt.

The time delay between sending a prompt and receiving a response. Important for real-time applications and user experience.

A neural network trained on vast amounts of text data to understand and generate human language. Examples include GPT-4, Claude, Llama, and Gemini.

The ability of models to process very large amounts of text in a single prompt. Long-context models can handle entire documents, codebases, or book-length inputs.

An efficient fine-tuning technique that trains only a small number of additional parameters rather than the full model. Makes fine-tuning more accessible and cost-effective.

The trained AI system that processes inputs and generates outputs. Different models have different capabilities, sizes, and specializations.

AI systems that can process and generate multiple types of content: text, images, audio, and video. GPT-4V and Claude 3 are examples of multi-modal models.

Human language as spoken or written naturally, as opposed to programming languages or formal notation. LLMs are designed to understand and generate natural language.

The field of AI focused on enabling computers to understand, interpret, and generate human language. LLMs represent a major advancement in NLP.

AI models whose weights and architecture are publicly available for anyone to use, modify, and build upon. Llama and Mistral are examples of open-source LLMs.

The structure and style of the model response. Can be specified in prompts to get responses as lists, tables, JSON, markdown, or other formats.

The learned values (weights) in a neural network that determine its behavior. Model size is often described by parameter count (e.g., "70 billion parameters").

A defined character or role for the AI to adopt when responding. Setting a persona helps create consistent, contextually appropriate responses.

The input text you provide to an AI model to get a response. The quality and structure of your prompt significantly affects the quality of the output.

The practice of crafting effective prompts to get desired outputs from AI models. Involves techniques like few-shot learning, chain-of-thought, and structured formatting.

A security vulnerability where malicious input tricks an AI into ignoring its instructions or revealing sensitive information. Similar to SQL injection for databases.

A technique that combines language models with external knowledge retrieval. The model first searches a database for relevant information, then uses it to generate more accurate responses.

A training technique where models learn from human preferences and ratings. RLHF helps align models to be more helpful, harmless, and honest.

A prompting technique where you assign the AI a specific role or expertise to adopt. Helps generate more contextually appropriate and expert-level responses.

The process of selecting the next token during text generation. Different sampling strategies (temperature, top-p) affect randomness and creativity.

A number that initializes the random number generator for reproducible outputs. Using the same seed with the same prompt should produce identical results.

Receiving AI responses token by token as they are generated, rather than waiting for the complete response. Improves perceived responsiveness.

AI responses in a specific, parseable format like JSON, XML, or markdown tables. Useful for programmatic processing of model outputs.

Instructions provided to the model that set its behavior, personality, and constraints. System prompts are typically hidden from end users and persist across conversations.

A setting that controls randomness in AI outputs. Lower values (0-0.3) produce more focused, deterministic responses. Higher values (0.7-1.0) produce more creative, varied outputs.

The basic unit that language models use to process text. Roughly equivalent to 3/4 of a word in English. Both prompts and responses are measured in tokens.

The algorithm that converts text into tokens for a language model. Different models use different tokenizers, affecting how text is split and processed.

The ability of AI models to call external functions or APIs to accomplish tasks. Enables models to search the web, run code, access databases, and more.

A sampling parameter that limits token selection to the smallest set of tokens whose cumulative probability exceeds p. An alternative to temperature for controlling randomness.

The text corpus used to train a language model. The quality, size, and composition of training data significantly impacts model capabilities and biases.

The neural network architecture underlying modern large language models. Introduced in 2017, transformers use attention mechanisms to process sequences of text.

A database optimized for storing and searching embeddings (vectors). Essential for RAG systems that need to find relevant content quickly.

The ability of AI models to understand and analyze images. Vision-capable models can describe images, answer questions about them, and extract information.

Asking a model to perform a task without providing any examples. Relies entirely on the model pre-trained knowledge and instruction-following abilities.