Glossary
43 terms. The language of AI-assisted development — from vibe coding to agentic workflows, context windows to MCP servers.
A
Agentic Coding
A development approach where AI operates as an autonomous agent — reading files, running commands, writing code, and iterating on errors — rather than just suggesting completions. The developer sets direction and reviews output while the AI handles execution across multiple steps.
Agents
In AI coding, agents are autonomous sub-processes that can be spawned to handle specific tasks — code review, testing, security scanning, architecture analysis. They operate independently with their own context and tool access, then report back results.
AGENTS.md
A configuration file used in Claude Code to define custom agents (subagents) that can be spawned during a session. Each agent definition specifies a name, model, tools, and a system prompt. Placing agent files in your project's .claude/agents/ directory makes them available as specialized workers for tasks like code review, testing, or architecture analysis.
AI Code Review
Using an AI model to review code changes for bugs, security issues, performance problems, and style violations. Can be triggered manually or automated as part of a CI/CD pipeline or agentic workflow.
AI-Assisted Development
The broad category of using AI tools to help write, debug, refactor, or understand code. Ranges from inline autocomplete (Copilot-style) to fully autonomous agents (agentic coding). Most developers today sit somewhere on this spectrum.
Autonomous Coding
A mode where AI works independently on a task with minimal human intervention. The developer defines the goal and constraints, then the AI plans, implements, tests, and iterates until done. Requires high trust in the AI's judgment and strong guardrails.
C
CI/CD Pipeline
Continuous Integration / Continuous Deployment. An automated workflow that builds, tests, and deploys code whenever changes are pushed to a repository. In agentic coding, CI/CD pipelines can trigger AI agents for automated code review, run tests the AI wrote, or deploy changes the AI made — all without manual intervention.
Claude Code
Anthropic's agentic coding tool. Available as a CLI that runs in your terminal, as an extension for VS Code and JetBrains IDEs, and as a web interface. Has full access to your filesystem, shell, and development tools. Operates as an agentic coding assistant that can read, write, execute, and iterate autonomously.
CLAUDE.md
A configuration file that gives Claude Code persistent instructions about your codebase — coding style, architecture decisions, common patterns, and project-specific rules. A default CLAUDE.md is created when you install Claude Code, but the real power comes from creating project-specific ones in each project root. You can also nest them in subdirectories for more granular control. Automatically loaded at the start of every session.
Code Generation
The act of AI producing source code from natural language descriptions, existing code context, or structured specifications. Quality depends on the model itself, how much context it has, and how specific the instructions are.
Context Rot
The gradual degradation of an AI coding session's effectiveness as the conversation grows longer and older context gets compressed or lost. As the context window fills up, the model loses track of earlier decisions, file states, and instructions. Symptoms include the AI repeating mistakes it already fixed, forgetting project conventions, or contradicting its own earlier output. Managed by keeping sessions focused, using persistent configuration (CLAUDE.md), and starting fresh sessions for new tasks.
Context Window
The maximum amount of text (measured in tokens) that an AI model can process in a single conversation. Larger context windows let the model "see" more of your codebase at once. Claude's context window is among the largest available, which matters for understanding complex codebases.
Copilot
GitHub's AI coding assistant. Supports multiple models including those from OpenAI and Anthropic. Primarily known for inline autocomplete inside your editor, suggesting code as you type. Has expanded into chat and agentic features but originated as the first mainstream AI-assisted development tool.
Cursor
An AI-first code editor built as a fork of VS Code. Integrates AI deeply into the editing experience with features like multi-file editing, codebase-aware chat, and inline code generation. Takes an IDE-centric approach to agentic coding, though tools like Claude Code can also be used inside Cursor or VS Code as extensions.
D
Diff
A representation of changes between two versions of a file, showing what was added, removed, or modified. In AI coding, diffs are how you review what the AI changed. Engineers read diffs line by line. Non-engineers can focus on testing the outcome and keeping track of high-level architectural decisions instead — what matters is that you verify the AI's work, however that looks for your skill level.
E
Extended Thinking
A mode where the AI model allocates extra internal reasoning tokens before responding. Produces better results for complex architectural decisions, debugging, and multi-step planning at the cost of longer response times.
G
Git
A distributed version control system that tracks changes to files over time. The backbone of modern software development — it lets you save snapshots of your code (commits), work on features in parallel (branches), and merge changes together. AI coding tools rely heavily on git to track what changed and revert mistakes.
GitHub
The most widely used platform for hosting git repositories, collaborating on code, and managing projects. Alternatives include GitLab, Bitbucket, and Gitea. In agentic coding, these platforms integrate with CI/CD pipelines, code review bots, and AI agents that can open pull requests, run checks, and respond to issues.
Guardrails
Constraints and safety mechanisms that limit what an AI agent can do. Examples include file permission rules, command allowlists, mandatory human approval for destructive operations, and automated testing gates. Essential for safe autonomous coding.
H
Hallucination
When an AI model generates plausible-sounding but incorrect information — inventing APIs that don't exist, referencing non-existent files, or fabricating function signatures. In coding, hallucinations are caught by actually running the code, which is one reason agentic tools (which execute and iterate) outperform pure suggestion tools.
Harness
The runtime environment that wraps an AI model and gives it access to tools, files, and system commands. In Claude Code, the harness is what connects Claude (the model) to your terminal, filesystem, and configured tools. It manages permissions, executes hooks, enforces guardrails, and handles the conversation loop. The model reasons; the harness acts.
Headless CMS
A content management system that provides content through an API without a built-in frontend. Unlike traditional CMS platforms (WordPress, Webflow), a headless CMS separates content storage from presentation, letting developers build the frontend with any framework. Examples include Sanity, Contentful, Strapi, and Directus.
Headless Mode
Running an AI coding agent without interactive human oversight — typically via a script, CI pipeline, or scheduled task. The agent operates fully autonomously within predefined guardrails. Useful for automated code review, test generation, and maintenance tasks.
Hooks
Event-driven automation in Claude Code. Shell commands that execute automatically before or after specific tool calls. For example: auto-formatting code after every file edit, running linters after writing, blocking destructive operations like force-pushes or file deletions, or validating changes against project rules. Defined in your settings and run without manual intervention.
Human-in-the-Loop
A workflow where a human reviews and approves AI actions at key decision points rather than letting the AI operate fully autonomously. The default and recommended mode for most agentic coding — you set direction, the AI executes, you review the results.
L
LLM (Large Language Model)
The type of AI model that powers coding assistants like Claude, GPT, and Gemini. Trained on vast amounts of text and code to predict and generate language. The "large" refers to the billions of parameters that encode the model's knowledge.
M
MCP (Model Context Protocol)
An open protocol that lets AI models connect to external tools and data sources — databases, APIs, file systems, browsers, and more. In Claude Code, MCP servers extend what the AI can do beyond reading files and running shell commands.
P
Plan Mode
A feature in AI coding tools where the AI creates a structured implementation plan before writing any code. Available in Claude Code, Cursor, and other agentic tools. Useful for complex features where you want to align on the approach before execution begins. The plan can be reviewed, modified, and approved before the AI starts building.
Prompt Engineering
The practice of crafting effective instructions for AI models. In agentic coding, this extends beyond single prompts to system-level configuration (CLAUDE.md), skill definitions, and structured specifications that shape how the AI approaches your entire project.
R
RAG (Retrieval-Augmented Generation)
A technique where the AI retrieves relevant information from external sources (documentation, codebases, databases) before generating a response. Reduces hallucinations by grounding the AI's output in actual data rather than relying solely on training knowledge.
S
Scaffolding
Using AI to generate the initial structure of a project, feature, or component — boilerplate code, file organization, configuration files, and basic implementations that you then build on. Agentic tools excel at scaffolding because they can create multiple files and verify the structure actually works.
Shell
A program that interprets and executes text commands on your computer. Common shells include bash, zsh, and PowerShell. When an AI coding tool "runs a command," it's executing that command in a shell. The shell is what gives agentic tools their power — it's the interface between the AI and your entire development environment.
Skills
Reusable workflow commands in Claude Code (invoked with /<name>) that encode multi-step processes — planning features, running reviews, deploying code, creating content. They give the AI deep procedural knowledge about how you work, turning complex workflows into single commands.
Spec-Driven Development
A methodology where you write a specification (what to build, acceptance criteria, constraints) before the AI writes any code. The spec becomes the contract: plan against it, execute from it, verify with it. Prevents scope drift and gives the AI clear success criteria.
Subagents
Agents spawned by a parent agent to handle specific subtasks in parallel — one for code review, another for security scanning, a third for testing. Each subagent gets its own context and tool access, reports results back, and is cleaned up when done. In Claude Code, you can define your own custom subagents in AGENTS.md files to encode specialized roles and workflows for your project.
T
Terminal
The application that provides a text-based interface to your computer's shell. On macOS it's Terminal.app or iTerm, on Windows it's Windows Terminal or Command Prompt. CLI-based AI tools like Claude Code run inside a terminal, giving them the same access to your system that you have when typing commands.
Tokens
The fundamental units that AI models use to process text. A token is roughly 3-4 characters or about 0.75 words. Token counts determine both the cost of API calls and how much text fits in a model's context window. Code tends to use more tokens per line than prose.
Tool Use
The ability of an AI model to call external functions — reading files, running commands, searching the web, editing code. This is what separates agentic AI from chat-only AI. The model decides which tools to use and in what order to accomplish a goal.
V
Vercel
A cloud platform for deploying and hosting web applications, especially popular with frontend frameworks like Next.js, SvelteKit, and Nuxt. Connects to your GitHub repository and automatically builds and deploys your site on every push. Provides edge functions, analytics, and preview deployments for pull requests.
Vibe Coding
A term coined by Andrej Karpathy describing a casual approach to coding with AI where you describe what you want in natural language and let the AI generate the code, often without fully understanding or reviewing the output. Emphasizes speed and experimentation over control and precision.
VS Code
Visual Studio Code. A free, open-source code editor by Microsoft and the most popular editor among developers. Supports extensions that add AI capabilities — including Claude Code, GitHub Copilot, and others. Many AI-first editors like Cursor and Windsurf are built as forks of VS Code, inheriting its extension ecosystem.
W
Windsurf
An AI-powered code editor (formerly Codeium) that combines inline completions with agentic capabilities. Positions itself between Copilot's autocomplete approach and Claude Code's full terminal autonomy.
Z
Zed
A high-performance code editor built in Rust, designed for speed and collaborative editing. Supports AI-assisted features including inline generation and chat. Differentiates from VS Code-based editors with its native performance and built-in multiplayer editing for real-time pair programming.
Follow the work
I write about building with AI as I figure it out. Tools, workflows, honest takes. No spam, no pitch.