Artifacts with AI, connected apps and data

By the end, you will distinguish a static interface, a Claude call, an app connection and storage. Example outputs are teaching material, not a promise of identical behavior in every account.

Not every Artifact needs AI. A button can reveal fixed feedback or calculate a sum without a model call. AI adds flexibility, latency, usage and possible mistakes. An app connection may read or change external information. Storage may retain data between visits. Add only what the task needs.

Create a fictional course-update editing Artifact.
The user pastes a draft. Show a fixed checklist:
Is there an unsupported date? Is a proposal described as approved?
Are numbers preserved? Are missing details marked?
At this stage, do not call Claude, connect services or store information.
State clearly that nothing is sent to another person.

Verify that it shows a checklist, not a claim to have checked facts. An interface can look intelligent without using AI. Clear labels avoid misleading the user.

Use fictional text: “48 registered; a choice is needed between two online-session cohorts.” No email address or personal information is needed.

Documentation checked on September 30, 2026 describes Artifacts calling Claude on Anthropic infrastructure. Users sign in and usage counts against their own limits; the exercise needs no pasted API key. New Artifacts request permission on first use of Claude.

Add a Check wording with Claude button.
Send the draft and source text entered by the user.
Return unsupported claims, a corrected draft and missing details.
Without a source, do not claim facts were verified.
Show waiting, success and error states.
Do not add sending, service connections or storage.

Read the permission request. Check exactly what goes to the model and what happens with an empty source. If unavailable, keep the static version and do not label fixed text as model-generated output.

Source: Unit 3 is being edited. No completion date was provided.
Draft: Unit 3 is ready and we will launch next week.

Illustrative correct output:

Unsupported claims: the unit is ready; launch is next week.
Corrected draft: Unit 3 is still being edited; no completion date was provided.
Missing: an approved completion date and an approved launch date.

Also test an accurate draft, empty source and long input. A failure state must say the check did not finish, rather than displaying stale output as a new analysis.

If results are wrong, revise the call instructions and presentation. The user still compares against the source. A button is not an approval stamp.

On eligible plans, Artifacts can use connected apps, including reading and writing. Initial approval presents apps and tools; some can be disabled. The documentation says choices may also apply to later uses.

Each shared-Artifact user connects their own apps rather than inheriting yours. Tools requiring approval for every action are unavailable to Artifacts according to the current source. Do not assume every later action will ask again.

Before a real connection, ask: is read-only enough? Is writing unnecessary? Which account is connected? What happens without access? Do not connect Slack or a calendar just to demonstrate a button when fictional data works.

The documentation describes storage on Pro, Max, Team and Enterprise on Web and Desktop. Personal storage holds each user's own private data. Shared storage exposes common information to users of the Artifact. The builder chooses where data goes; a field named “personal note” does not establish privacy.

Start without storage. For storage practice, save only fictional progress:

Store only the user's personal exercise-completion flag.
Do not store source text or drafts.
Show what is retained and provide a way to delete the test flag.
No shared storage or other users' data.

Close and reopen to test persistence. Legacy Artifacts require publication for storage; new ones do not according to the documentation. Do not publish merely to test storage without choosing an audience. The source describes a 20MB text-only limit; this is not a general file drive or database for every task.

The full sequence: simple version, assess AI need, approve a defined call, test accurate and contradictory input, inspect storage, then consider sharing. Each layer needs its own test.

Easy: static or AI

Build the fixed checklist without AI. Success: it labels itself a checklist, not automatic fact verification.

Intermediate: source versus draft

Add AI and test the Unit 3 contradiction. Success: both unsupported claims are marked and the correction preserves the source.

Challenging: information boundaries

Write down what goes to the model, what is stored and who can access it. Success: each has an explicit answer rather than an assumption that everything in an Artifact is always private.

Another user cannot run it: check their login, eligibility and connections. Do not share your key or private data as a workaround.

Stale result after an error: clear previous output and display a direct failure message.

Storage fails: check new versus legacy behavior and storage policy. Do not publish automatically.

Unexpected information visible: stop real input and inspect personal versus shared storage and audience settings.

Local export behaves differently: exported files may not include AI, connections or storage. Test separately.

Fast app creation does not remove the need for permissions, error handling and result verification.