AI Glossary for Designers: Key Terms Explained in One Line

Which terms explain how a model works?

• LLM: a large language model, trained on lots of text to predict the next token.
• Token: the unit of text a model processes: a word, a fragment or a symbol. Size and cost are measured in tokens.
• Context window: the maximum number of tokens a model handles in one request, including its own answer.
• Embedding: a list of numbers representing a text or image, so that similar things end up close together.
• Knowledge cutoff: the date up to which training data reaches; anything later is unknown to the model on its own.
• Hallucination: a false answer delivered with confidence, an effect of how text is generated and not an intentional lie.
• Reasoning model: a model that produces intermediate steps before answering, at the price of more time and cost.
• Multimodal model: a model that processes more than one kind of data, such as text, images, audio or video.
• Temperature and top-p: parameters that control how varied or conservative the choice of the next token is.
• Open vs. closed model: weights available to host, as opposed to a model reachable only through an API.

How do you give a model instructions and data?

• Prompt engineering: designing instructions, context, examples and format to get useful, consistent results.
• System prompt: permanent instructions that define the model’s role, tone and limits.
• Few-shot prompting: including a few input and output examples so the model follows the pattern.
• Chain of thought: getting the model to write intermediate steps before the final answer.
• Structured output: a response in a fixed format, usually JSON, that an application can read without interpreting free text.
• Context engineering: deciding what goes into the model at each step, not only the prompt.
• RAG: retrieving relevant passages from documents and adding them to the request so the answer is grounded in them.
• Fine-tuning vs. prompting: further training the model on your data, as opposed to guiding it with instructions in each request.
• Prompt caching: reusing parts of the prompt already processed to lower cost and latency.

Which terms describe an agent?

• AI agent: a system where a model receives a goal, decides the steps and uses tools until it finishes.
• Assistant, copilot and agent: the first responds, the second suggests inside your task, the third runs several steps.
• Tool use: the model asks to run an external function; the application runs it and returns the result.
• Agent loop: the cycle repeated until done: the model decides, a tool runs and the result comes back.
• MCP: an open standard for connecting AI applications to files, databases and tools.
• Autonomy levels: a five-level scale (Feng, McDonald and Zhang) based on the role the person takes toward the agent.
• Human-in-the-loop: a person approves or corrects at the points where an error costs the most.
• Plan mode: the agent investigates and proposes changes, but does not act until you approve the plan.
• Subagent: a specialized agent with its own context and permissions that returns only the final result.
• Skill: a folder with a SKILL.md that teaches the agent a procedure and loads only when needed.
• Sandboxing and permissions: operating-system isolation plus rules on which tools need approval.
• Computer use: the model observes the screen and operates mouse and keyboard as a person would.

How do you measure and control quality?

• Evals: tests for an AI system; the equivalent of automated tests for behavior that varies.
• Golden dataset: test cases with reference answers, used to compare versions in a repeatable way.
• LLM as judge: a model grades answers against a written criterion; it is only reliable if you validate it with people.
• Benchmark: a standardized test for comparing models; the score only describes what the test measures.
• Prompt regression: a change that makes cases that used to work fail.
• Guardrails: controls that check inputs, outputs or actions and can block them; they do not guarantee total safety.
• Red teaming: attacking the system on purpose to find failures before someone else does.
• Model bias: a repeatable tendency that gives worse or unfair results for certain groups.

Which terms matter when designing the experience and the product?

• Streaming: showing the answer in pieces as it is generated; the model is not faster, the text just appears sooner.
• Trust calibration: making the person’s trust match the system’s real reliability.
• Uncertainty: communicating how sure the system is and how to verify what it says.
• Explainability: giving the user what they need to know where a result comes from and when to check it.
• Human handoff: passing the conversation to a person with enough context that nobody repeats everything.
• AI dark patterns: designs or dialogue that push the user to believe or do something they would not have chosen.
• Generative UI: an interface where the model decides which components to show, instead of them being fixed in advance.
• AI-first vs. AI-added: a product designed around AI from the start, as opposed to AI added to an existing one.
• AEO: a marketing term for preparing a site so AI answer engines cite it; it is not a standard.
• llms.txt: a proposed Markdown file with a site summary and key links for AI agents.
• Vibe coding: building software by describing it to an AI and accepting the code without reviewing it closely.
• Design engineer: a profile between design and frontend that prototypes and builds in code; the role has no standard definition.
• Design tokens: named design values in a standard format, which an agent can use without guessing a hex.
• Code Connect: a Figma tool that links a Figma component to its code component.

Example

Illustrative example: in a meeting someone says “the agent hallucinated because it ran out of context window”. With this glossary you can separate the terms. The context window is the token limit per request. If the agent lost its initial instructions, the problem is context engineering. If it invented a fact, that is a hallucination, and the defenses are RAG with citations, or an eval that catches it. Three different problems that sounded like one in the meeting.

Common mistakes

• Using “agent” for any AI chat. An assistant that responds and an agent that runs several steps with tools call for very different permission and supervision designs.
• Believing temperature 0 guarantees the same answer. Anthropic’s documentation states that even then results are not fully deterministic.
• Confusing RAG with fine-tuning. RAG adds documents to the request; fine-tuning changes the behavior of the model itself.
• Treating a benchmark as a verdict. The score describes performance on that test, not on your product; for that you need your own evals.
• Using AEO as a guarantee. It is a marketing term, not a feature with its own settings.

Sources

• Glossary (Claude documentation)
• What is the Model Context Protocol (MCP)? (modelcontextprotocol.io)
• Building effective agents (Anthropic)
• Levels of Autonomy for AI Agents (Feng, McDonald and Zhang, arXiv:2506.12469)
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al., arXiv:2005.11401)
• The /llms.txt file (llmstxt.org)