Artifacts: documents and apps Claude builds beside the chat

What an artifact is, and why it's separate

An artifact is a piece of content Claude builds in a dedicated panel to the right of your chat, instead of inside the conversation. When what Claude is producing is substantial and self-contained (roughly more than a dozen lines, and something you'd want to keep rather than just read), it opens an artifact for it.

Why does that matter? Because content in the chat scrolls away and content in an artifact stays put. The panel is a workbench. You can edit the thing directly, ask Claude to change it and watch it update in place, flip back to an earlier version, and copy or download the finished item. A summary buried twelve messages up is a chore to find and reuse; the same summary as an artifact sits open beside you, editable, until you're done with it.

Artifacts come in a few shapes:

• Documents: a policy draft, a briefing, meeting minutes, a plan, in formatted text.
• Tables and structured content: an action tracker, a comparison grid, a checklist.
• Diagrams and visuals: a simple flowchart or org diagram.
• Interactive tools and apps: a small working thing, such as a calculator, a form, or a single-page mini-app Claude builds for you.
• AI-powered artifacts: an app that has Claude's intelligence built into it, so it can do smart things when other people use it. More on these below.

Artifacts run on all plans. The panel is enabled by a capability setting (look under Settings → Capabilities for the toggle that covers code execution and file creation), so if you never see a side panel, that's the first thing to check. The official artifacts help article has the current path.

The full loop on a Fernway task

Here's the whole cycle (build, edit, version, reuse) on something real. Maya needs an action tracker from a messy set of meeting notes.

From the meeting notes below, build an action tracker as a table with four columns: Action, Owner, Due date, Status. Mark any action with no clear owner as UNASSIGNED and set Status to 'To do' for all of them. Notes: [paste the Fernway meeting notes].

Claude opens an artifact containing the table. It's now sitting in the panel, not lost in the chat. Two ways to change it:

Ask Claude. Type an instruction in the chat and the artifact updates in place:

Move all the UNASSIGNED rows to the top so they're impossible to miss, and add a fifth column, 'Priority', with High, Medium or Low based on the due dates.

Or edit it directly. For a text or Markdown artifact you can click into the panel and change wording yourself (fix a name, tweak a date) without asking Claude at all.

Each time the artifact changes, Claude keeps the previous versions. There's a way to step back through them (a version control at the edge of the panel), so if an edit made things worse you can return to the one that worked. Your own manual edits don't overwrite Claude's memory of the original, so you can keep iterating from a known-good point.

When it's right, get it out: the panel gives you copy (straight to your clipboard, to paste into an email or doc), download (as a file), and, for things you want others to see, publish, which creates a shareable version. Publishing is a deliberate step: only publish content you're happy to make visible, and never anything confidential.

Now, from that same tracker, write a three-sentence status update Maya could paste into an email to Priya, just the headline of where things stand and what's still unassigned. Keep the table as it is.

AI-powered artifacts, briefly

Some artifacts don't just hold content. They use Claude when someone interacts with them. You can describe an app in plain English ("a tool where a new starter answers a few questions and gets a personalised first-week checklist") and Claude builds a small app with its own intelligence inside. When a colleague uses it, it responds smartly to what they type. These are powerful for prototyping an idea without writing code, and they work across the plans. They're also the point where you should slow down: an AI-powered artifact can be shared and used by others, so think about what it might say to them before you publish. The dedicated guide, Prototype AI-powered apps with Claude artifacts, covers what's possible.

Try it now

Common mistakes

• Copying content in and out of the chat instead of editing the panel. The whole point of an artifact is that it stays put and updates in place. Ask Claude to change it, or edit it directly. Don't paste it into the message box and back.
• Losing a good version by over-editing. If a change makes it worse, use the version control to step back rather than trying to undo by hand. Earlier versions are kept for exactly this.
• Publishing without thinking. Publish makes an artifact shareable. Never publish anything with confidential names, figures or client detail, and remember an AI-powered artifact will respond to whoever uses it.
• Trusting a slick artifact more than a plain reply. A formatted table or a working little app looks authoritative, and a tidy table of the wrong dates is still wrong. The polish is presentation, not proof. Check the contents of an artifact against your source material exactly as you'd check any Claude answer, especially figures, dates and owners before anyone acts on them.

Keeping current

Artifacts are one of Claude's fastest-moving areas: new types, richer interactivity and better sharing arrive often. The core loop (build in the panel, edit, version, copy out) is durable; the specifics will grow. Check the official artifacts help article and the Claude release notes when something looks new. Accurate as of 13 July 2026.