AGENTS.md is a project instruction file that gives AI coding agents task context, workflow rules, and constraints so agent behavior aligns with team standards.

Agent Skills are modular capability definitions that package domain-specific guidance, tools, and patterns for recurring agent tasks.

Hooks are automation points tied to lifecycle events that let teams enforce checks, trigger scripts, or modify behavior around agent actions.

Approval Modes are policy settings that control which operations an agent can run automatically and which require explicit human approval.

Automations are recurring or event-based executions that let AI agents perform routine work automatically under predefined instructions.

A System Prompt is a top-level instruction layer that sets goals, boundaries, and response style before user messages are processed.

An agent branch is a separate working version of a project where AI-made changes can be reviewed before they affect the main site.

An agent branch protects the live project by isolating AI edits. Teams can preview the result, compare changes, request revisions, and merge only when the work is ready.

An agent prompt is a user request written to guide an AI agent toward a specific project outcome, such as a page edit or CMS update.

An effective agent prompt gives the AI enough direction to act without guessing. It usually names the target, the desired outcome, constraints, and any content or style rules that should be preserved.

An agent guardrail is a boundary that prevents unsafe, unwanted, or inconsistent AI behavior during project work.

Agent guardrails can stop destructive edits, prevent duplicate CMS entries, preserve brand rules, or require review before publishing. They are especially useful when agents work across large websites or shared team projects.

Conversational editing is the process of changing a website through natural-language feedback instead of direct manual edits alone.

Conversational editing lets a user describe what should change and have an agent update the project. It works best when each request names the target clearly and the agent keeps the result editable for follow-up refinements.