1. LandOffer
  2. Blog
  3. All articles
  4. Before Sharing a Claude Artifact at Work: Check Viewers, Connected Apps, and Shared Storage

AI at Work

Before Sharing a Claude Artifact at Work: Check Viewers, Connected Apps, and Shared Storage

Separate artifact viewers, connected operations and shared storage before a work pilot. Includes a filled fictional review and explicit not-run acceptance script.

9 min readLandOffer team

Before sharing a Claude artifact at work, check three separate boundaries: who can open it, which actions it can perform through each user's connected apps, and which information it saves for other users to see. A restricted source connection does not make copied information private once the artifact puts it into shared storage. Confirm the artifact generation, account plan, intended readers, and data paths before choosing a sharing route.

This LandOffer workplace guide reviews Anthropic's public documentation checked on October 25, 2026. It includes a fictional ticket-triage tool and a completed review of its proposed test script. No Claude account, artifact, connector, invitation, or storage behavior was tested or changed for this article. The example's expected results are acceptance requirements, not observed product outcomes.

Colleagues considering separate doors, personal work trays and a shared display before opening a workshop to visitors

Identify which artifact you are about to share

Record where and when the artifact was made, its current version, and the sharing controls visible in your account. Do not reuse an old tutorial merely because its screenshot resembles your tool.

Anthropic's artifact overview calls chat artifacts made before September 16, 2026 legacy artifacts. Earlier Cowork live artifacts are another category. Identify the applicable documentation before deciding how to distribute a work tool; a new artifact's instructions do not establish an older one's behavior.

The sharing guide describes private starting access and different routes by plan and artifact type. New shared artifacts generally require a Claude account; published legacy chat artifacts have an exception. Legacy sharing also deserves an attachment review: the guide warns that sharing a chat artifact can expose files in that chat. Publishing and then unpublishing a legacy artifact is not a harmless test: the documented unpublish operation permanently deletes its stored data and prevents republishing that artifact.

For a work review, capture the relevant distinction in ordinary language: “New internal artifact, not an older published chat tool; no original customer attachments included.” If the generation or source conversation is uncertain, keep distribution on hold until the owner can establish it. Do not publish just to discover which storage instructions apply.

Keep opening, operating, and saved information separate

Anthropic describes connected-app access as using the viewer's own connections; missing source access produces an error rather than the creator's source data. Its sharing guide also distinguishes restricted invitations from broad link access. Artifacts using connected apps or Claude cannot use the broad anyone-with-link route. Outside invitees have additional functional limits. Check the actual artifact's available route instead of assuming that an external invitation is equivalent to an internal coworker's access.

The overview says approved connected tools may read or write, and users can disable individual tools. It distinguishes personal storage from shared storage, where users work with the same saved information. The builder chooses which data goes where; a first-use shared-storage confirmation is documented. These documented features do not establish which fields your implementation reads, copies, displays, or saves.

The practical review question is where information goes after a permitted read. A tool might retrieve a synthetic ticket using one tester's connection, then save its text to a common triage board. Another tester could read that saved copy without independently retrieving the original ticket. That possibility follows from the proposed design; do not interpret source authorization as a complete storage-access policy.

Inspect rendered content, saved fields, comments, exports, and diagnostic messages as distinct disclosure paths. If you cannot explain a field's audience, omit it from the proposed pilot so an uncertain data path cannot become a routine shared field. “The connector is restricted” does not answer who can read a copied description in the shared board.

A completed access and storage review for a fictional tool

The fictional tool, Queue Lantern revision 3, suggests a routing category for synthetic office-equipment tickets. Builder Keiko proposes a limited internal pilot on an approved Team workspace. Internal testers Sol and Fen have different test-source permissions. External colleague Bastian is outside the organization and is not an authorized reader. These people, accounts, permissions, and artifact settings are invented planning inputs.

The review uses exactly three synthetic tickets: S-101, desk-lamp replacement; S-102, meeting-room marker refill; S-103, spare keyboard request. There are no names, customer records, confidential incidents, screenshots, credentials, or actual workplace identifiers. Routing is a suggestion for human review, not an action that creates or changes a real ticket.

Boundary or field Revision 3 review finding Required disposition before any pilot
Open the artifact Keiko proposes named internal testers only; organization-wide and external distribution are outside the pilot Confirm the selected reader list and workspace identity; a copied link alone is not evidence of the intended audience
Edit the artifact Keiko maintains the reviewed version; Sol and Fen need to use the tool, not alter its implementation Verify the actual available access levels support the intended use; do not assume editing rights are necessary for interaction
Read test-source tickets Sol's approved test source includes S-101 and S-102; Fen's includes only S-103 Use each tester's own approved connection; a denied read stays denied, with no creator-account fallback
Change the source system The prototype originally offered a button to create a routing task Remove that action from the proposed pilot and exclude write tools; no live-ticket mutation is required
Personal scratch text Draft reasoning stays outside the common-board design Inspect implementation and verify isolation with synthetic markers; the word “personal” in a label is insufficient evidence
Shared board Original design saved full ticket descriptions; revision 3 proposes only synthetic ticket ID, suggested category, and review state Explicitly treat every saved board entry as visible to participating readers; do not copy restricted source text into it
Diagnostics and exports An error template included the attempted ticket description Replace it with a minimal failure notice and non-sensitive test reference; inspect export content separately

This is a filled design review, not an account screenshot or a security certification. The review finds an actual difference within the fictional versions: the earlier full-description board would have defeated the intended separation between Sol's source access and Fen's. The selected revision removes that copying path from the proposed design. Whether the implementation does so remains a test gate.

Keiko also records what the pilot cannot establish. It cannot prove general isolation across all inputs, validate every future connector, authorize real employee data, or approve outside sharing. The work sponsor, Daria, can approve the limited workflow exercise; the appropriate organizational owner still controls account permissions and data-use approval. Keiko cannot grant those approvals by documenting a preference.

Treat organization settings as another review layer

Anthropic's Team and Enterprise admin guide separates broad external-link sharing from email invitations outside the organization. Turning one on does not enable the other. It also notes individual artifact exceptions to external-link restrictions. Therefore a reviewer should inspect the effective artifact access, not infer it solely from an organization-wide switch.

The guide further distinguishes enabling artifact connectors from choosing individual apps: the organization control is not described as an app-by-app allowlist. Its capability controls and sharing settings are different layers. Turning Artifacts off does not revoke previously shared links, according to the guide. Do not use a feature toggle as a substitute for checking existing access.

In Queue Lantern's review record, Daria asks the organizational owner to confirm only the settings relevant to the proposed internal pilot. The record does not ask them to enable outside sharing, broaden source permissions, or bypass a disabled connector. If the required approved configuration is unavailable, the alternative is a synthetic static walkthrough, clearly described as noninteractive. A missing feature is a reason to narrow the exercise, not to move company data into a personal account.

A workshop review separates one person's material from the shared display and keeps an unapproved visitor outside the working area

A finished test script with expected results, not invented passes

The following is Queue Lantern's completed preflight specification. Each step has a named tester, a precise input, an expected boundary, and a decision rule. No step was executed while writing this article. The result column is deliberately recorded as “not run”; the script defines the evidence still needed before anyone can record a product test as passed.

Case and authorized setup Input and action to perform only after approval Expected result and release decision Current result
Sol, internal approved test reader with S-101 and S-102 test-source access Open the reviewed revision through the approved internal route; request S-101 Correct synthetic source content may appear; unavailable tools or unexpected broader access halt the case for investigation Not run
Fen, separate internal test reader with only S-103 source access Request S-101, then S-103, without adopting Sol's account or connection S-101 must not reveal its source content; S-103 is the allowed comparison. An access error is not fixed by sharing the builder's credentials Not run
Sol then Fen, approved shared-board participants Sol saves S-101's synthetic ID, category and review state; Fen opens the board Fen can see the intentionally shared minimal entry, but not the source description. Any extra copied field blocks the pilot Not run
Sol then Fen, personal-scratch isolation check Sol enters marker LANTERN-SOL-ONLY in the proposed private scratch area; Fen checks their own area and offered exports Fen must not receive Sol's marker. Absence in one view is insufficient if exports or another saved path disclose it Not run
Bastian, outside reader with no invitation or pilot authorization Organization-approved reviewer evaluates the uninvited-reader case with synthetic content only; do not send a real work link to an unauthorized colleague Artifact content must remain unavailable under this pilot's intended access. Unexpected access blocks distribution; do not broaden permissions to make the case easier Not run
Keiko, version and action-scope check Inspect the reviewed revision and offered tool approvals before either tester proceeds No task-creation action or unnecessary write tool is accepted. If revision or tool list differs, repeat the design review Not run

These expected outcomes combine documented access principles with the pilot's stricter design requirements. They do not predict the exact wording of an error screen. Record the actual account identity, revision, visible permission choice, synthetic inputs, and outcome when an authorized tester eventually runs a case. Redact any evidence before circulating it.

The external-reader case is a boundary to verify through an approved test arrangement. It is not permission to send a company artifact to Bastian. If the organization cannot authorize a suitable negative-access test, record that coverage as missing and hold broader distribution. A private browser window alone does not represent every signed-in outside identity or invitation state.

Decide what is ready and what remains blocked

The completed review decision is to retain Queue Lantern as a synthetic preflight candidate. Keiko owns implementation inspection and removal of unnecessary write behavior. Sol and Fen are proposed testers, subject to the organizational approval route. Daria owns the workflow decision after evidence is available. No one has accepted a real-data rollout or external audience.

Daria's next review requires separate results for source permissions, shared fields, personal scratch, exports, and unintended-reader access for revision 3. A successful open is not enough. For a failed check, record the exposed boundary, affected synthetic input, proposed correction, and version to retest before reconsidering distribution. Do not mark the entire artifact safe because one synthetic ticket behaved as expected.

Before distributing your own work artifact, save the intended audience, reviewed version, connected operations, personal and shared fields, outstanding tests, and authorized decision owner. Keep real data out until the applicable checks and organizational approvals are complete. Return to the current official instructions when the artifact generation or configuration changes. More workplace guides are available through LandOffer.

Sources and Further Reading