Artificial Intelligence Glossary

The "brain" behind Claude, ChatGPT or Gemini. It's trained on enormous amounts of text to predict the next word — and from that seemingly simple skill emerge conversation, reasoning and coding.

The instruction you give the AI. Prompt quality determines answer quality: it's the number-one skill of any AI user. We have a whole library of proven prompts.

The discipline of designing effective prompts: giving context and a role, showing examples, asking for step-by-step reasoning and defining the output format. The difference between knowing it and not is 10x in results.

The smallest unit of text a model processes (roughly ¾ of an English word). API prices are measured in tokens, and so is the AI's "memory" (its context window).

How much information the AI can "hold in mind" at once: the conversation, attached documents, repository code. Measured in tokens. Bigger window = it can work on larger projects without forgetting the beginning.

An AI that doesn't just answer: it plans, uses tools, executes actions and completes tasks end to end with minimal supervision. Claude Code and Cowork are agents. The word of 2026.

Open standard created by Anthropic that connects AI to external tools (Gmail, GitHub, databases…). It's the "USB-C of AI": any service publishes an MCP server and any compatible AI can use it. Full guide in Connectors.

A folder of instructions and resources that teaches the AI to do a specific task at expert level (build a perfect Excel, follow your company's process). It activates on its own when the task matches. Learn to build your own.

An installable package bundling skills, connectors and commands into one piece. Plugin marketplaces let you install complete capabilities with a single command.

A technique that lets the AI consult external documents (your knowledge base, your PDFs) before answering, instead of relying only on what it memorized during training. It reduces errors and enables working with private, up-to-date data.

When the AI confidently states something false: it invents facts, quotes or code functions that don't exist. Mitigated with RAG, good instructions ("if you don't know, say so") and always verifying critical information.

Partially retraining a model with your own data to specialize it. Today, for most cases, skills and RAG achieve the same with far less cost and complexity.

A numerical representation of a text's meaning. It lets machines "measure" how similar two sentences are — the foundation of semantic search and RAG.

Using an already-trained model to generate an answer. When you chat with Claude, you're doing inference. Its computational cost is what APIs charge per token.

An AI that understands and/or generates several content types: text, images, audio, video. The frontier models of 2026 are all multimodal to varying degrees.

The model's ability to "think before answering": spending extra compute analyzing the problem step by step. Extended-reasoning models shine at math, complex code and planning.

A parameter controlling answer randomness: low = more predictable and precise; high = more creative and varied. Adjustable in APIs; calibrated for you in consumer apps.

A standardized exam to compare models (math, code, knowledge…). Useful as reference, but beware: the best AI for you is the one that best solves YOUR tasks, not the one winning the latest benchmark. We explain this in the comparison.

Models whose parameters are published so anyone can run them on their own infrastructure (Llama, Mistral…). The counterpoint to "closed" models like Claude or GPT, used via API or apps.

A technique to try to bypass an AI's safety protections and make it generate content it would normally refuse. Labs invest ever more in resisting them — even with evaluation frameworks shared between competitors.

A training method created by Anthropic: the model learns from an explicit set of written principles (its "constitution") in addition to human feedback. It's part of Claude's identity. More curiosities here.