AI coding glossary

A plain-English glossary of the terms you will actually run into while building with AI in 2026. Skimmable, bookmarkable, linkable. If a term is missing, it probably is not worth knowing yet.

Agent

An AI model equipped with tools and a loop — it can read files, run commands, observe results, and decide what to do next. The difference between a chatbot and an agent is that an agent acts.

Agentic loop

The iterative cycle of observe → think → act → observe that an AI agent runs to complete a task. Modern coding agents loop until the task is done or a budget is exhausted.

Chain of thought

A prompting technique where the model is asked to reason step-by-step before answering. Often produces more reliable output on complex tasks.

Completion

The model's response to a prompt. In coding tools, an inline completion is a suggested continuation of the line you are typing.

Context window

The amount of text a model can hold in attention at once, measured in tokens. A larger context window lets the model reason over more code at the same time, but adding irrelevant content dilutes attention.

Embedding

A vector representation of a piece of text or code. Embeddings let systems compare meaning — used to retrieve the most relevant files or docs for a query.

Fine-tuning

Adjusting a pre-trained model's weights on new data. Rarely necessary for vibe coding in 2026 — prompting and context are almost always enough.

Foundation model

A large, general-purpose model trained on broad data that can be adapted to many tasks. GPT-5, Claude Opus 4.6, and Gemini 3 are foundation models.

Grounding

Giving the model authoritative source material (docs, code, data) so its output is anchored to facts instead of its prior.

Hallucination

When a model generates plausible-sounding but false information — an API that does not exist, a parameter with the wrong name. The main failure mode to guard against.

Inference

Running a trained model to generate output. Every prompt you send is an inference call.

In-context learning

The ability of large models to adapt to new tasks from examples in the prompt alone, without any training. Why few-shot prompting works.

Latency

Time from prompt to response. Matters enormously in interactive coding — a 3-second completion feels great, a 30-second one breaks flow.

LLM

Large language model. The class of models vibe coding relies on.

MCP

Model Context Protocol. An open standard from Anthropic for connecting AI assistants to external tools and data sources. Increasingly the common interface between models and the outside world.

Multi-modal

A model that can process more than one type of input — text, images, audio, video. Useful in coding for feeding screenshots, designs, or diagrams into a prompt.

Prompt

The input you send to a model. In coding, good prompts have intent, constraints, and relevant context.

Prompt engineering

The craft of structuring prompts to get reliable, useful output. Real — but often overstated. Clear thinking beats clever prompts.

RAG

Retrieval-augmented generation. A pattern where the system retrieves relevant documents or code first, then includes them in the prompt. How most AI IDEs answer questions about your codebase.

Reasoning model

A model trained or configured to spend extra inference time "thinking" before answering, producing better results on hard tasks at the cost of latency.

Rules file

A project-level file (.cursorrules, AGENTS.md, CLAUDE.md) that encodes conventions the AI must follow on every prompt. The single highest-leverage piece of context you can set.

Sampling

How the model picks the next token during generation. Controlled by parameters like temperature and top-p.

System prompt

A prompt that sets the model's persona and constraints for an entire session, before any user message. Used by tools to enforce defaults.

Temperature

A sampling parameter from 0 to 2 that controls randomness. 0 means deterministic, higher means more varied. For coding, low temperatures are almost always better.

Token

The unit a model reads and writes in. Roughly 3–4 characters in English. Pricing and context limits are measured in tokens.

Tool use

The ability of a model to call external functions — read a file, search the web, run a shell command. The feature that turns a language model into an agent.

Vibe coding

Building software primarily by collaborating with an AI model. See the full definition on our what is vibe coding page.

Vector database

A database optimized for embedding search. Used by coding tools to find the most relevant snippets for your query.

WebContainer

Browser-based runtime (from StackBlitz) that runs full Node.js stacks without a server. What powers Bolt and similar in-browser IDEs.

Zero-shot / few-shot

Zero-shot: asking the model to perform a task with no examples. Few-shot: including a handful of examples in the prompt. Few-shot prompting remains one of the most reliable accuracy wins.

Keep learning

If a term here sparked a question, the long-form guides cover the concepts in context: what is vibe coding, how to start, the workflow, and the tools.