Status: Supplemental research, not a normative parity source. Current public behavior is
grounded in the official docs and specified in docs/product-spec.md.
Undocumented binary strings may describe gated, experimental, or removed behavior.
Distilled from the current official Artifact guide and linked references, the historical
launch post, and a focused string survey of the locally installed Claude Code 2.1.233 binary
(skill registry, tool registration, and consent machinery). Installation, source coverage,
and authentication limits are recorded in
docs/evidence/claude-code-host-verification.md.
The official contract and binary observations are deliberately separated below. The local host was not signed in, so no claude.ai publish, share, connector, or administration flow was executed.
- A permissioned
Artifacttool publishes.html,.htm, or.mdas one live page at a private claude.ai URL; every publish is a version and open viewers update in place. - Sharing supports public links and, on Team/Enterprise, organization viewers and editors. Public availability and administration depend on plan and organization policy.
- A declared claude.ai MCP connector runs through each viewer's connection and approval. Connector-backed artifacts cannot be public, and local MCP servers are capture-time only.
- Pages have no application backend, use a strict CSP, and are limited to 16 MiB rendered.
- Users can disable the tool through settings, environment, or permission rules. Organization controls cover enablement, connectors, public sharing, role scopes, retention, audit, and Compliance API lifecycle operations.
| Skill | Role | Load discipline |
|---|---|---|
artifact-design |
design fundamentals, anti-cliché list, naming rules | mandatory before writing any artifact |
artifact-capabilities |
window.claude.* runtime roster |
gate: must load before declaring capabilities |
artifact-diagramming |
diagram craft, inline SVG, mermaid | when drawing |
artifact-pr-review |
PR review pages | when reviewing |
workshop |
decide-and-revise loop pages | when workshopping |
artifact-dashboard |
KPI tiles + primary chart + breakdown table, slot template | dashboards |
artifact-data-table |
sortable/filterable table, slot template | tabular datasets |
artifact-explainer |
step-by-step teaching pages, slot template | explainers/tutorials |
artifact-components |
embeddable reusable components with hash-pinned scripts | when a page carries decisions |
plan-artifact |
publish the implementation plan as a page | plan review |
adjacent: dataviz, whiteboard/canvas, prototype |
chart honesty method (with a runnable palette validator), canvas board, app prototype | referenced from artifact-design |
The template skills share a discipline worth copying: slot markers that must ALL be replaced,
"never invent a value", provenance footers, and number-formatting rules — all now in our
docs/component-spec.md and skill reference.
Fully absorbed: artifact-design, artifact-dashboard, artifact-data-table, artifact-explainer,
artifact-pr-review (structure), workshop, plan-artifact, artifact-components (philosophy).
Not applicable: artifact-capabilities (no window.claude.* runtime).
Absorbed in skills/artifact-pages/reference/visuals.md: artifact-diagramming (mechanism-over-name, labeled arrows, one-figure-one-claim) and dataviz (form heuristic, color-by-meaning,
anti-pattern list, interaction discipline). Their runnable palette validator stays unported —
our palettes are fixed token sets, so there is nothing to validate.
Plus: plan-artifact flow (offer to review a plan as a page) and a reusable-component skill (first entry: the workshop decision component).
The single public Artifact tool contains internal action names for publish, live-edit
(gated), list, read, watch, unwatch, status, comments, resolve, page data, decisions, and a
small database. Their presence is not evidence that every action is enabled or supported for
every account.
- Publishing: same-URL redeploy by file path, versions, restore, share menu with version picker, gallery, description subtitle, emoji favicon, title scan (first 8KB), auto-open, Ctrl+] reopen
- Collaboration: comments (span-quoted threads, resolve, analyst subagent pipeline), editor role (Team/Enterprise), shared per-artifact database
- Runtime: capabilities (live data / shared state / downloads / self-update), MCP connectors called as the viewer's account at view time
- Guards: publish permission prompt, fail-closed consent floors, stale-version guard with autoread recovery, sensitive-delta analysis for live-shared republishes
- Admin: org toggle, RBAC scoping, retention policies, compliance API, public-sharing toggle, connector toggle
Adopt: chart honesty ("title the finding, not the axes"), fail-closed consent, layered guards, naming-like-a-product, copy-as-design-material, token-cost discipline (SVG over raster, drop unneeded interactivity, summarize datasets).
Invert (with justification): fixed renderer + Markdown/JSON authoring instead of hand-written per-page HTML (consistency, diffability, safety, token cost); single-file inline instead of multi-file publish; Markdown as the preferred format (their warning assumed Markdown = no craft; our renderer supplies the craft).
See docs/claude-code-comparison.md for the parity matrix and evidence.