The language of AI
Enough vocabulary to follow the conversation — models, tokens, prompts, agents, and the ways AI can be wrong.
Models, LLMs, and reasoning models
Artificial intelligence (AI) is the broad category: software that performs tasks associated with human intelligence.
An AI model is a trained mathematical system that recognises and generates patterns. GPT, Claude, and Gemini refer to families of models; ChatGPT, Claude, and Gemini are also names used for products built around models. The product may add memory, search, files, tools, and a user interface.
A large language model (LLM) processes and generates language — including programming languages — by predicting likely continuations from the context it receives.
A reasoning model is optimised to spend more computation planning and checking difficult tasks. It can be useful for architecture, complex debugging, and multi-step analysis. A faster general model is often enough for copy edits, visual variations, or small code changes.
Tokens and context windows
A token is a chunk of information processed by a model. A word may be one token or several. Code, images, and files also consume context in model-specific ways.
The context window is the amount of information available to the model for a response. It may include your messages, earlier replies, files, project rules, selected code, and tool results.
Multimodal AI
A multimodal model can work across more than text — for example images, audio, PDFs, screenshots, or video. For designers, this means you can provide evidence directly:
- a screenshot of a broken layout;
- a Figma frame showing hierarchy and spacing;
- a research report to synthesise;
- a content model or tone-of-voice guide;
- a screen recording showing an interaction bug.
Prompts, instructions, and examples
A prompt is the immediate request you give the model.
Persistent instructions are standing rules that apply across many tasks — for example coding conventions, design-system usage, writing style, or validation commands.
Few-shot examples are a small set of examples showing the desired pattern. They are especially valuable for tone of voice, naming, structured outputs, and recurring component patterns.
Agents, tools, skills, and MCP
An agent can plan and take actions to pursue a goal. It may inspect files, edit code, run a browser, or query another system.
A tool is one action the agent can use, such as reading a file or running a test.
A skill is a reusable set of instructions or a workflow for a particular job, such as reviewing accessibility or preparing a usability-test plan.
Hallucinations and verification
A hallucination is a plausible-sounding output that is unsupported or false. In coding, this might be a package that does not exist, an invented API method, or a confident explanation that does not match the actual code.
Use the right kind of verification:
- factual claims → check a reliable source;
- package or framework behaviour → check official documentation;
- code changes → run the app, tests, and build;
- visual fidelity → compare the rendered result at relevant sizes;
- user claims → trace them to research evidence;
- accessibility → combine automated checks with keyboard and human review.
- What is the difference between a model and a product built around a model?
- Why should important decisions live in project documentation rather than only in a long chat?
- Give one example of hallucination in your discipline.