Glossary

The words used across the workshops, in plain language. Where a workshop teaches a word, the entry links to it.

---

Agent. An AI model given a role, instructions and sometimes tools, so it acts in a particular way. Workshop 2

AGENTS.md. A file of standing instructions that coding assistants read at the start of every session. Workshop 5

API. Short for application programming interface: a way for one program to ask another for something. Your prototype uses an AI provider's API to send context to a model and get text back. Workshop 4

API key. A secret code that lets your app use a paid service, such as an AI model. Anyone who has it can spend your money, so keep it out of the browser and out of GitHub. Workshop 4

Attention. The part of a transformer that lets each token take the other tokens in the context into account when the model predicts what comes next. Workshop 1

Bodystorming. Prototyping an interaction by acting it out with people, some playing the user and some playing the system. Workshop 2

Branch. A separate line of work in Git, so changes don't affect your main version until you merge them. Workshop 5

Case study. The story of one project, told for someone who hasn't seen it: what you made, how you made it, and what it taught you. Workshop 6

Chatbot. A program you talk to in a conversation. Many AI chatbots are a language model with a chat window around it. Workshop 3

Claude Code. A coding assistant made by Anthropic. It can read your project, change files and run tools for you. Workshop 5

Codex. A coding assistant made by OpenAI. It does the same job as Claude Code. Workshop 5

Coding assistant. An AI tool that works inside your project: it proposes plans, writes and changes code, and runs commands while you review and steer. Workshop 5

Commit. A saved snapshot of your project's files in Git, with a short message about what changed. Workshop 5

Completion. The text a language model produces to continue the text it was given. Workshop 1

Context. Everything the model is given to predict from: the system instructions, the knowledge, the conversation so far and the user's latest message. Workshop 1

Context window. The most context a model can take in at once, measured in tokens. Anything beyond it is left out. Workshop 1

Deployment. One version of your project put online, so others can use it. Vercel makes a new deployment each time you push to GitHub. Workshop 4

Design surfaces. The places where an AI product's design lives: system instructions, parameters, the model, knowledge and tools. Workshop 1

Embedding. A list of numbers that represents the meaning of a piece of text, so a computer can find texts with similar meanings.

Environment variable. A setting kept outside your code, such as an API key. On Vercel, you add them in your project's settings. Workshop 4

Evaluation (eval). A test that gives your prototype a set input and compares its response with what you expected. Testing Your AI Prototype

Fine-tuning. Further training of a model that has already been trained, on a smaller set of examples, to change how it behaves. Workshop 2

Foundation model. A large model trained on a broad range of data, which other products and models are built on. Workshop 2

Function calling. A way for a model to ask your program to run a specific function, such as a search or a calculation, and use the result. See also tool use.

Git. A program that tracks changes to a project's files, so you can save snapshots and go back to them. Workshop 5

GitHub. A website that stores Git projects online, so you can share them and connect them to services like Vercel. Workshop 4

Hallucination. A model's response that sounds confident but is false or made up. It happens because the model predicts likely text, not true text.

Knowledge base. Reference material you give an agent to draw on, such as documents or notes. Workshop 1

LLM. Short for large language model: a model trained on large amounts of text to predict the next token. Workshop 1

Model. The trained system that makes the predictions. Different models give different results for the same context, so choosing one is a design decision. Workshop 1

Neon. A service that hosts databases online. The toolkit uses it when a prototype needs to remember things. Workshop 5

Open-weights model. A model whose trained numbers, its weights, are published, so anyone can download and run it on their own computer. See Going Further.

Parameters. Two meanings. In a model's settings, parameters such as temperature change how it responds. Inside a model, parameters are the billions of numbers learned during training. Workshop 1

Playwright. A free tool from Microsoft that lets a coding assistant open your prototype in a browser, click through it and take screenshots. Workshop 5

Pretraining. The first and largest stage of training, where a model learns to predict the next token from a huge amount of text. Workshop 2

Prompt. A message you send to an AI model or tool. Workshop 1

Push. Sending your commits to GitHub. For projects connected to Vercel, pushing is how you publish. Workshop 5

RAG. Short for retrieval-augmented generation: finding relevant pieces of a knowledge base and adding them to the context before the model responds.

Rate limit. The most requests a service lets you make in a period of time. Past it, requests fail until the limit resets. Workshop 4

Reasoning model. A model trained to work through a problem in steps before it gives its answer. Workshop 2

Repository. A project tracked by Git. Your project on GitHub is a repository. Workshop 4

RLHF. Short for reinforcement learning from human feedback: a training stage where people rate a model's responses, and the model is adjusted toward the ones they preferred. Workshop 2

Role card. A one-page description of an agent's role: its purpose, rules, boundaries, what it doesn't do, its knowledge, and its inputs and outputs. Workshop 2

Sampling. How a model picks the next token from the probabilities it predicted. It doesn't always pick the most likely one. Workshop 1

System card. A role card kept as a versioned record of your agent's design. Documenting an AI Design

System instructions. Text that tells the model its role and rules before any user message. Workshop 1

System prompt. Another name for system instructions.

Temperature. A setting that controls how much randomness goes into sampling. Low temperature picks likely tokens; high temperature allows less likely ones. Workshop 1

Token. A piece of text a model reads and writes: often part of a word, sometimes a whole word or a punctuation mark. Workshop 1

Tokenizer. The program that splits text into tokens. Different models use different tokenizers. Workshop 1

Tool use. An agent's ability to use tools outside the model, such as search, code or other programs. Workshop 1

Transformer. The kind of neural network most language models are built on. Attention is its key part.

Vector store. A database of embeddings, used to find the pieces of text most similar in meaning to a question.

Vercel. A service that publishes websites and web apps from a GitHub repository. Workshop 4

Vercel Blob. Vercel's service for storing files, such as images people upload. Workshop 5

Vibe coding. Building software by describing what you want to an AI tool and steering it by the results, rather than writing the code yourself. Workshop 3

Vocabulary. The full set of tokens a model knows. The model can only read and write text made of those tokens. Workshop 1

VS Code. Visual Studio Code, a free code editor from Microsoft. The toolkit's coding assistants run inside it. Workshop 5