Artifacts: turning a conversation into something you can use

Somewhere in a conversation the output stops being text you read and becomes a thing you use. That is an artifact: "Artifacts (opens in new tab) are interactive apps that emerge from your conversations with Claude": a calculator, a diagram, a small web tool, a chart, rendered in a panel beside the chat. The remarkable part is real; the discipline that keeps it useful is the same discipline the rest of this site teaches, pointed at a new surface.

The loop, and why it starts with the problem

Anthropic's own workflow puts the problem before the artefact and expects iteration from the first render; the support article walks the mechanics. What matters here is why that ordering works. Leading with "build me a dashboard" gets you a generic dashboard; leading with the problem — what the thing is for, who uses it, what "done" looks like — is the right-context rule doing its usual work, and concrete examples beat adjectives here too: paste the actual numbers, name the real audience. The first render is a draft, exactly as answers vary means any first output is a draft. The conversation is the editor.

The trap the docs name out loud

The single most important sentence in the support article is a warning, and it is this site's whole thesis in miniature: "Artifacts look finished as soon as they render, so spend two minutes trying to break yours before anyone else sees it". A rendered app carries the same false authority as a fluent paragraph: polish is not correctness, and here the polish is a working UI.

The article is blunt about the stakes: "An app that looks polished but gets the numbers wrong will mislead the people you share it with, and you are the only one who can catch that before they do". A calculator that computes the wrong total, a chart that mislabels an axis, a form that drops an edge case: all render beautifully. This is verifying what you cannot judge applied to your own creation: put in inputs whose answers you already know, and check the artefact against them before it goes anywhere.

Sharing turns a private draft into a published thing

The reach is a few clicks: "Share your finished artifacts with others by clicking the Publish button to generate a link", and "Anyone with the link can use and interact with your artifact without signing in", with one carve-out: "AI-powered artifacts are the exception: those require sign-in". Two consequences worth holding before you click:

Publish is publication. A link anyone can open is a thing you have put into the world, with your unchecked numbers in it. The break-it-first step is not optional once other people can reach it.

What you put in it, you shared. An artefact built from a spreadsheet you pasted carries that data. Before publishing anything built on real information, the what-never-goes-in question applies: would you email this content to a stranger? A public link is closer to that than it feels.

What goes wrong

Trusting the render. It looks like a finished product the instant it appears, and looking finished is precisely what it is good at. Test it.

Leading with the format, not the problem. "Make me an app" produces a generic one. The interview the docs describe exists to extract the specifics; skip it and you get the average.

Publishing the unchecked calculator. The wrong total, shared, believed, acted on, and you were the only one positioned to catch it.

Pasting sensitive data into a thing you then publish. The convenience of "build it from my real data" and the exposure of a public link are one careless click apart.

One-shot expectations. Abandoning a near-miss instead of saying "close, now change this." The iteration is the tool, not a sign the first try failed.

How to check it worked

Before sharing any artefact, run three known inputs through it — cases where you already know the right answer — and confirm it agrees. For anything with numbers, that is the difference between a useful tool and a confidently-wrong one wearing a nice interface. If you cannot construct three checks, you do not yet understand the artefact well enough to publish it, and that is the finding.