Claude Code Glossary: Skills, Hooks and Artifacts Explained

An artifact is a deliverable the AI produces outside the conversation, editable and shareable. A skill is a procedure the agent learns once and reuses across sessions. A hook is an automatic rule that fires at a precise point in the agent's work, without human intervention. Here are the three terms explained, with business examples, and their rough equivalents across the main tools on the market.

What are an artifact, a skill and a hook, in one sentence each?

An artifact is content generated by the agent and displayed separately from the chat thread, which you can review, correct and share without copying it out of the conversation. A skill is a procedure written once, in a format the agent recognises and reuses from one conversation to the next. A hook is a rule that runs automatically at a precise moment, before or after an agent action, without a human needing to think about it.

These three concepts solve three different problems: the deliverable that gets lost in a long conversation, the procedure that needs re-explaining every time, and the rule that gets forgotten once in ten. They often combine in a single project.

What is an artifact, in practice, and why does it matter for a business?

In an ordinary conversation, everything scrolls past in the same thread: questions, answers, successive drafts. An artifact takes the document out of the thread and places it in a separate space, editable without losing context, then exportable or shareable through a link.

For a business, this changes how you work on a real deliverable. A business owner having a brief or a landing page drafted gets a self-contained document, editable in small steps ("shorten paragraph two", "add a price column"), rather than a block of text to copy out again on every iteration. An artifact can also be a visualisation: a dashboard, a page mock-up shown to a client. Worth remembering: it is not a file stored permanently by default, but a working object that has to be explicitly saved to keep it.

What is a skill, and why is it not just a copy-pasted prompt?

A prompt, however well written, lives inside a single conversation. For the AI to apply the same method the following week, someone has to remember it, find it or rewrite it. A skill removes that friction: a procedure saved once, which the agent applies automatically whenever the situation calls for it.

Technically, a skill takes the form of one or more instruction files, with metadata describing when to use it. The agent reads that description, judges whether the skill fits the current request, then applies it. A concrete example: a restaurant business that issues weekly invoices with specific legal wording writes an "issue a compliant invoice" skill once, with the expected format. The agent then applies it to every invoice. This very article was produced from a skill describing the site's editorial position. For a step-by-step method that does not require writing code, see how to create a Claude Code skill without being a developer.

What is a hook, and why should a non-technical business owner know the word?

A hook is a rule that fires automatically at a precise event in the agent's work: before it uses a tool, after it, or at the end of a session. Unlike an instruction in a prompt, which an agent could in theory forget as a conversation gets longer, a hook runs every time.

It is the difference between politely asking someone to check a document before sending it, and installing a control that blocks the send until the check has happened. A hook moves a rule from "goodwill" to "automatic safeguard".

Useful business examples: a hook that refuses to save a file containing an API key into a shared folder, a hook that reformats a document to house style before it is presented, or a hook that checks no raw personal data is written into a report meant for a third party. This connects directly to what we cover in our full guide to AI agents: a useful agent is bounded by rules that apply without exception.

How does this show up in Claude and Claude Code?

In Claude's conversational products, the artifact is the most visible mechanism: a document, a page or a visualisation the AI generates opens in a separate panel, editable through iterations and shareable via a link.

In Claude Code, the command-line tool built for software development, the balance is different. The agent works directly in the project's files rather than in a separate panel, so the deliverable takes the form of a modified file on disk, which makes the artifact less central. Skills and hooks, on the other hand, are formalised there: a skill is declared through a file with metadata, invoked according to the task at hand; a hook is configured explicitly, following Claude Code's official hooks documentation.

And how does it translate to ChatGPT or Codex?

ChatGPT offers a separate editing panel, Canvas, close to the artifact idea: a document or a snippet of code opens there, apart from the thread, and gets reworked in small edits, even though the implementation differs from Claude's.

For skills, the equivalence is actually direct: Codex adopts the same open SKILL.md format as Claude Code (a standard published by Anthropic and documented at agentskills.io), a folder with an instruction file and metadata, invokable individually depending on the task. ChatGPT, on its side, offers custom GPTs with their own instructions and knowledge files, a single-context-per-project logic closer to a persistent prompt than a packaged skill. For the hook in the strict sense, public formalisation on OpenAI's side remains lighter: it is worth checking the tool's documentation before relying on it.

And on Antigravity, Google's agentic IDE?

Antigravity is Google's agentic development environment, more recent than Claude Code or Codex. The agent produces work plans and proposed changes displayed in a dedicated interface, which is close to the artifact logic: a structured result, distinct from a simple exchange of messages, reviewable before it is applied.

On skills, Antigravity also adopts the open SKILL.md format (stored in .agents/skills/ at project or global level), the same standard as Claude Code and Codex: a skill written for one works on the other without adaptation. On hooks, however, Antigravity has its own format (hooks.json, with its own lifecycle events), distinct from Claude Code's even though the underlying idea (triggering a rule automatically at a precise moment) is identical.

Should you expect an exact match between these tools?

No, and that is the most important point. These three concepts describe real needs (a separate deliverable, a reusable procedure, an automatic rule) that exist, in one form or another, in most serious agentic tools, but the vocabulary and the level of formalisation vary between vendors, and change quickly.

For a business choosing a tool, the useful question is not "does it have exactly an artifact, a skill and a hook like Claude Code", but "does it let me produce a reusable deliverable, lock in a recurring procedure, and enforce a rule that applies without exception". That is the logic behind our guide to building and deploying an AI agent in your business and the full definition of an AI agent. Before writing any code with the help of these tools, it is also worth getting comfortable with the technical basics: databases, OAuth, API keys.