AI coding glossary

43 terms from the world of AI coding tools, defined in plain English with what each one actually means when you are prompting.

Jump to: Agents, Modes, Context, Files and config, Prompting, Practices, Building, Evaluation, Tokens and cost

Agents 10 terms

Agent loop: The agent loop is the repeating cycle an AI agent runs: decide what to do next, take an action such as reading a file or running a command, observe the result, and repeat. It ends when the task is complete, a limit is reached, or the agent asks for input.
Agent skills: Agent skills are reusable packages of instructions, context, and sometimes code that an agent loads when a task calls for them.
Agentic coding: Agentic coding is using an AI tool that plans and executes a whole task in your codebase rather than completing one line at a time. The agent reads files, makes edits across several of them, runs commands, checks the result, and iterates until the task is done or it gets stuck.
Autonomous coding agent: An autonomous coding agent completes a described task without step-by-step supervision: it plans, edits, runs, and verifies on its own, reporting back when done. Autonomy is a spectrum, and the practical question is how many decisions it makes before you next see the work.
CLI coding agent: A CLI coding agent is a coding agent you invoke from the command line. It works on the repository in your current directory, uses your real toolchain, and returns changes as edits and a summary rather than as suggestions in an editor.
Coding agent: A coding agent is an AI system that carries out programming tasks end to end: reading a codebase, editing multiple files, running commands, and reacting to the output. It differs from a code assistant in that it decides its own sequence of steps rather than responding to one request at a time.
Headless agent: A headless agent runs without an interactive interface — invoked by a script, a CI job, or a scheduler rather than a person typing. It takes a task as input, works autonomously, and produces output such as a diff, a pull request, or a report.
Subagent: A subagent is an agent spawned by another agent to handle a scoped piece of work in its own context window.
Terminal AI agent: A terminal AI agent runs in your shell rather than inside an editor. Because it has the same access you do, it can install dependencies, run migrations, drive git, and execute your test suite, which lets it take on tasks that editor-based tools cannot complete on their own.
Tool calling: Tool calling is the mechanism that lets a model do things beyond producing text. The model emits a structured request naming a tool and its arguments, the surrounding system executes it, and the result comes back as input for the model's next step.

Modes 5 terms

Agent mode: Agent mode is the setting in an AI coding tool where the model can act rather than only answer: creating and editing files, running terminal commands, and iterating on the result. It contrasts with a chat or ask mode, which explains and suggests without touching your code.
Ask mode: Ask mode is the read-only setting in an AI coding tool: the model answers questions about your code and explains what it finds, but does not edit files or run commands. It is the safe default for understanding a codebase before changing it.
Composer mode: Composer is Cursor's multi-file editing interface, where you describe a change and the tool proposes coordinated edits across several files at once, presented as diffs you review together rather than one file at a time.
Plan mode: Plan mode makes an AI coding tool propose an approach before it edits anything.
YOLO mode: YOLO mode is the informal name for running an AI agent with approvals disabled, letting it edit files and run commands without asking each time. It is much faster and removes the review step that would otherwise catch a bad change before it lands.

Context 5 terms

Codebase indexing: Codebase indexing is the process an AI editor uses to build a searchable representation of your repository, usually by splitting files into chunks and storing vector embeddings so it can retrieve code by meaning rather than exact text match.
Context engineering: Context engineering is the practice of deciding what information an AI model sees for a given task — which files, which conventions, which prior results — and in what form. It has largely replaced prompt phrasing as the thing that most affects output quality in coding tools.
Context retrieval: Context retrieval is how an AI coding tool decides which parts of your codebase to read for a given request. Approaches range from you attaching files explicitly, to semantic search over an index, to an agent running grep and reading what it finds.
Context window: The context window is the maximum amount of text a model can consider at once, measured in tokens and covering everything in the request: system prompt, your instructions, the files supplied, prior conversation, and the response being generated.
Repo map: A repo map is a compressed outline of a codebase — its files, key symbols, and how they relate — given to a model so it understands the shape of a project without reading every file. It buys structural awareness at a fraction of the token cost.

Files and config 4 terms

AGENTS.md: AGENTS.md is a tool-neutral convention for a Markdown file in a repository that gives AI coding agents the project context they need — build commands, conventions, architecture, and constraints.
CLAUDE.md: CLAUDE.md is a Markdown file in your repository that Claude Code reads at the start of every session.
Cursor rules: Cursor rules are project-level instructions stored in your repository that Cursor applies to every AI request — conventions, framework versions, patterns to follow, things never to do.
Custom instructions: Custom instructions are standing directions you give an AI tool that apply to every request rather than a single one — preferred language and framework, response style, conventions to follow. They can be set per user or per project depending on the tool.

Prompting 4 terms

Meta-prompting: Meta-prompting is using a model to write or improve a prompt rather than to do the task directly.
Prompt injection: Prompt injection is an attack where instructions hidden in content a model reads are treated as commands rather than data.
Prompt template: A prompt template is a reusable prompt with placeholders for the parts that change — the file, the framework, the requirement — so a structure that works can be applied to many tasks without being rewritten each time.
System prompt: A system prompt is the instruction set given to a model before any user message, defining its role, constraints, and behaviour for the whole conversation.

Practices 8 terms

AI code completion: AI code completion predicts and suggests the code you are about to write, usually as inline grey text you accept with Tab. It works from the current file, nearby open files, and recent edits, and operates at the scale of a line, a block, or a function.
AI code generation: AI code generation is producing working code from a natural-language description — a function, a component, a module, or an entire application. It spans everything from a chat request for one function to an app builder scaffolding a full project from a spec.
AI code review: AI code review is using a model to examine a diff or a file for bugs, security problems, and quality issues before a human reviews it. It works best as a first pass that catches mechanical problems, not as a replacement for review by someone who knows the system.
AI pair programming: AI pair programming is working alongside an AI assistant the way you would with a human pair: it suggests, you steer, and you review continuously rather than delegating a task and checking back later. The reviewing is the part that makes it pairing rather than outsourcing.
AI refactoring: AI refactoring is using an AI tool to restructure existing code without changing its behaviour — extracting functions, renaming concepts across a codebase, splitting large modules, or migrating between patterns.
PRD for AI: A PRD for AI is a product requirements document written to be consumed by an AI coding tool rather than by a team. It is shorter and more concrete than a traditional PRD, focusing on what must be built and constrained rather than on justification and background.
Spec-driven development: Spec-driven development means writing a precise specification before generating any code, then using that spec as the prompt.
Test-driven development with AI: Test-driven development with AI means writing or generating the tests first, then letting an agent implement until they pass.

Building 2 terms

Design to code: Design to code is generating working interface code from a visual input or a description of one — a screenshot, a Figma file, or written art direction. Output is typically React components with styling, ready to paste into a project.
Text to app: Text to app describes tools that generate a working application from a written description — pages, data model, and often auth and deployment — without you writing the scaffolding. The output is a running project with a live preview rather than a code snippet.

Evaluation 2 terms

LLM evals: LLM evals are systematic tests of a model or AI feature's output quality: a set of inputs, a definition of a good response, and a scoring method. They exist because you cannot tell from a handful of impressive examples whether something works reliably.
SWE-bench: SWE-bench is a benchmark that tests whether AI systems can resolve real GitHub issues from open-source Python repositories. A model is given an issue and the codebase, and its patch is scored by whether the project's own tests pass afterwards.

Tokens and cost 3 terms

Input tokens: Input tokens are the units of text a model reads: your prompt, the files supplied, the system prompt, and the conversation so far. They are billed separately from output and are usually much cheaper per token, though agentic tools consume them in far greater volume.
Output tokens: Output tokens are the text a model generates: explanations, code, and — with reasoning models — internal thinking that is billed even when it is not shown. They cost several times more per token than input, and they are capped per response.
Token limit: A token limit is a cap on how much text can be processed in a request or a period. It shows up in three forms: the context window's size, a maximum output length per response, and rate limits on tokens per minute or per month for an account.