Terminal
Terminal/Skill
Design guidance for the data screens of custom Databricks Apps built with AppKit and React: layout, charts, KPIs, and loading or error states.
By Databricks
Works with
npx skills add databricks/databricks-agent-skills --skill databricks-app-designMake Databricks data + AI apps that communicate clearly and compile to real AppKit code. This skill merges two bodies of knowledge and binds them to implementation:
references/dashboard-patterns.mdreferences/ibcs-notation.mdreferences/appkit-cheatsheet.mdFor package scripts and project-local CLI tools, use the parent skill's package-manager selection and command placeholders.
Design advice that doesn't name a real component is incomplete. Always end at a component plan.
databricks-aibi-dashboards), generic frontend (forms, auth, settings, marketing), or scaffolding/build/deploy (→ databricks-apps). A plain "create a dashboard" / "build a dashboard" request (no app / AppKit / React / custom-code signal) means a managed AI/BI (Lakeview) dashboard → use databricks-aibi-dashboards, not this skill. If a request is "add a form", "deploy this", or "build a Lakeview / AI-BI dashboard", this skill should not fire.databricks-apps builds/runs the app; this skill decides what the data screens should look like and which primitives realize them.dashboard-patterns.md (static / analytic / magazine / infographic / repository / embedded mini). State it.ibcs-notation.md rules: message-in-title, scenario marks (actual/PY/plan/forecast), honest scales, semantic color. On any chart-vocabulary conflict, IBCS wins (see the conflict note in that file).@databricks/appkit / @databricks/appkit-ui (see appkit-cheatsheet.md); never cite a component AppKit doesn't ship. There's no prebuilt KPI/trend/distribution card — compose those from primitives, following the notation rules. Use colorPalette + semantic tokens, never hardcoded hex. Bind data with useAnalyticsQuery/queryKey + sql.* params.Skeleton; Empty → Empty with a useful next action; Error → inline message, never a blank panel; Partial/stale → show what you have + a freshness note.Gate: this section applies only if the app has a Genie / chat / natural-language / "ask your data" surface. For a pure dashboard / KPI / report app with no conversational input, skip this section and references/genie-ai-trust.md entirely. When it does apply, implement ALL five (code in references/genie-ai-trust.md):
A Genie/chat/NL answer is only trustworthy if the user can see how it was produced and who it ran as. "Use GenieChat + a spinner" is NOT enough — for ANY Genie/chat surface, ship all five (copy the exact snippets from the reference):
/api/whoami route (real x-forwarded-email/x-forwarded-user headers) + the signed-in user in a Badge. Claim OBO only if user_api_scopes: [dashboards.genie] is wired; otherwise disclose the query runs as the app's service principal.attachments[].query in an inspectable "Generated SQL" Card; never hide how the answer was computed.useGenieChat().status (streaming/error), never a frozen spinner.genie() space config + a truthful execution-identity note (OBO when user-scoped, else service principal) + empty/error/ambiguous handling (Empty, Alert).Design proposal:
## Direction
[Genre, audience, primary task, design intent.]
## Pattern & notation choices
- Composition: [data info, meta info, layout, interaction, color]
- Notation: [message, scenario marks, scales, semantic color]
## Component plan ← the part that makes it buildable
- [element] → [AppKit component] (queryKey/props), [token/palette], states handled
## Tradeoffs & risks
[What's summarized/hidden/paginated/interactive; overload, scale, a11y, maintenance risks.]
Critique: lead with the top comprehension/integrity issue, cite the component/file, then list findings by impact, each with the concrete fix (which component/token/state to change).
KpiCard) — compose composites from published primitives instead.