Homeassistant Dashboard Designer

Home Assistant Dashboard Designer

Design dashboards as systems: predictable structure, reusable templates, minimal drift. Treat Stitch as inspiration only; translate to safe Lovelace YAML.

Primary Input: Natural Language (Default)

You can give natural-language guidance. The skill must infer the structured intent internally (dashboard intent, view name, entity mapping, constraints) while enforcing the design system and validating entities/services via the Home Assistant MCP when available.

Minimum helpful info to include in natural language:

• What to change (add/remove/refactor) and why (the goal).
• Where to change it: the exact view/partial path(s) under config/dashboards/**.
• Any constraints you care about: desktop/mobile columns, time window preferences for graphs, "do not touch" templates/sections.

If any of the above is missing, ask targeted clarifying questions (do not demand a full "intent block").

Example natural-language request:

• "Refactor config/dashboards/infrastructure/views/06_mariadb.yaml to match the button-card-first system used in other Infrastructure views. Keep the same entities, no new templates, 4 columns desktop / 2 columns mobile."

Optional Input: Structured Intent (Allowed, Not Required)

If the user provides structured intent, accept it. Do not require it.

dashboard_type: infra  # infra|home|energy|environment
view_name: "Infrastructure Overview"
target_paths:
  dashboard_yaml: "config/dashboards/infrastructure/dashboard.yaml"
  view_yaml: "config/dashboards/infrastructure/views/07_docker_containers.yaml"
entities:
  - name: "Proxmox01"
    cpu: "sensor.proxmox01_cpu"
    memory: "sensor.proxmox01_mem"
constraints:
  desktop_columns: 4
  mobile_columns: 2
notes:
  - "Preserve existing templates; minimize diff."

If updating an existing view and not obvious from context, ask:

• Which view file (single view dict) is the source of truth?
• Any "don't touch" areas/templates?
• Any fallback cards already in use that must remain?

Output Contract (Always Produce)

• One valid Lovelace view (single view dict)
• Deterministic ordering (stable ordering of cards/sections)
• Section comments (short, consistent)
• Zero per-card styling variance (no one-off styles; use templates/snippets)
• Fallback usage explicitly justified (YAML comments adjacent to the fallback card)

Card Priority Model

Attempt Tier 1 first. Use Tier 2 only if Tier 1 cannot express the requirement safely.

Tier 1 (primary):

• custom:button-card (structure + tiles + headers + containers)
• card-mod (shared/global polish only; shared snippets OK)
• custom:flex-horseshoe-card (single-metric radial telemetry)
• custom:mini-graph-card (single-metric trends)

Tier 2 (fallback, must justify in comments):

• entities, markdown, history-graph, conditional
• iframe (last resort)

Layout Rules (Hard Constraints)

• Allowed layout containers: grid, vertical-stack
• Maximum nesting depth: 2 (example: grid -> vertical-stack -> cards)
• No horizontal-stack inside grid cells (prefer: more grid columns, or vertical-stack)
• No freeform positioning
• No layout logic embedded in card-mod
• For dense tile lists in type: sections views, keep the outer panel full-width (column_span: 4) and make the inner tile grid responsive using custom:layout-card (4 desktop / 2 mobile unless the user asks otherwise).
• In type: sections views, the first section immediately below top chips/badges must be a full-width container (column_span: 4) with a single wrapper card that also sets grid_options.columns: full to prevent skinny-column whitespace.
• Consolidate upward after removals/moves: do not leave dead space, empty placeholders, or intentionally sparse rows unless the user explicitly asks for whitespace.

Note: If the repo/view uses Home Assistant type: sections, treat sections as the top-level structure and enforce the same container rules inside each section (sections should contain only grid/vertical-stack cards and their children).

Design System Rules (Implementation Discipline)

Button Card (Primary Renderer)

• Use custom:button-card as the structural primitive.
• Templates required. Do not inline per-card styles:; keep styling inside centralized templates/snippets.
• Logic via variables or template JS only (avoid duplicating logic per-card).

card-mod (Styling Constraint Layer)

• Allowed only for shared/global polish:
  • padding normalization
  • border radius control
  • shadow suppression
  • typography consistency
• No entity-specific styling
• No per-card visual experimentation

Flex Horseshoe Card

• One primary metric per card.
• Explicit color thresholds.
• No mixed units.

Mini Graph Card

• One metric per graph.
• Consistent time windows per dashboard.
• Minimal legends (only when required).

Template Architecture (Required)

• Views reference templates only (for custom:button-card: every instance must specify template:).
• Templates accept variables; views pass variables.
• Do not create new templates unless explicitly instructed.

Repo reality check (important):

• If the repo already has established templates (for example, Infrastructure uses bearstone_*), use what exists.
• If the spec's "required template categories" are not present, map to the closest existing template and call out the gap, or ask for explicit permission to add templates.

Stitch MCP (Inspiration Only)

Use Stitch only to learn:

• visual hierarchy (headers, grouping, density)
• spacing rhythm (padding/gaps)
• grouping patterns (panels, sections)

Never treat Stitch output as implementation code.

When prompting Stitch, include these constraints verbatim:

Grid-only layout. No absolute positioning. No JS frameworks. No custom HTML/CSS outside card-mod.
Card types limited to: custom:button-card, card-mod, custom:flex-horseshoe-card, custom:mini-graph-card, and HA core fallback cards only if unavoidable.
Max layout nesting depth: 2. No horizontal-stack inside grid cells.

Optional Stitch Handoff (Light Bridge)

Use this only when the user explicitly asks for visual ideation, layout exploration, or variants.

• Trigger Stitch path:
  • User asks for redesign inspiration, variants, or multiple visual directions.
• Skip Stitch path:
  • User asks for direct dashboard refactor/update and no ideation is needed.
• Preconditions:
  • Stitch is used for hierarchy/spacing/grouping ideation only.
  • Stitch output must never be copied as implementation code.
  • Final implementation must still obey all allowed card/layout rules in this skill.

Handoff artifact (summary only, no code export):

stitch_intent:
  layout_pattern: "<grid and grouping pattern to translate>"
  section_hierarchy: "<header/panel/section ordering>"
  density_target: "<compact|balanced|spacious>"
  cards_to_translate:
    - "<visual card concept mapped to allowed Lovelace card types>"

Translation requirement:

• Convert stitch_intent into approved Lovelace structures only (grid, vertical-stack, Tier 1 cards first).
• Re-map any unsupported Stitch concepts to compliant cards/containers and explain fallback choices inline when Tier 2 is required.

Fallback behavior when Stitch is unavailable:

• Continue with manual ideation using this skill's design system and rules.
• State degraded mode clearly ("Stitch unavailable; proceeding with manual hierarchy/density planning").
• Do not block dashboard work on Stitch availability.

Workflow (Do This In Order)

1. Use direct-update mode for this repo:
   • Edit production YAML directly (no staged dashboard copies in this skill workflow).
   • Keep rollback safety by minimizing diffs and validating before HA restart/reload.
2. Read the target dashboard/view/partials/templates and include targets before editing:
   • Confirm the exact source view/partial/template files.
   • Confirm referenced !include files/directories exist and are in-scope.
3. Determine intent from the user's request and existing dashboard context: infra (NOC), home, energy, or environment. Keep one intent per view.
4. Validate entities and services before editing:
   • Prefer the Home Assistant MCP for live entity/service validation (required when available).
   • Record the MCP validation step in the work notes before writing YAML.
   • If MCP is not available, ask the user to confirm entity IDs and services (do not guess).
5. Draft layout with constraints: a top-level grid and optional vertical-stack groups.
   • If using Stitch, first summarize stitch_intent and treat it as advisory input to this step.
   • After removals, reflow cards/sections upward to collapse gaps and reduce empty rows.
6. Implement using Tier 1 cards first; reuse existing templates; avoid one-off styles.
7. If fallback cards are necessary, add an inline comment explaining why Tier 1 cannot satisfy the requirement.
8. Validate in this order:
   • Run pwsh -NoProfile -File tools/validate_dashboards.ps1.
   • Run pwsh -NoProfile -File tools/ha_ui_smoke.ps1.
   • Run Home Assistant config validation (ha core check or homeassistant --script check_config) when available.
   • Run scripts/validate_lovelace_view.py from this skill against each changed view file.
9. Failure behavior:
   • If requirements can't be met: state the violated rule and propose a compliant alternative.
   • If validation fails: stop, surface the error output, and propose corrected YAML. Do not leave invalid config applied.

Example flow:

• User asks for redesign inspiration -> Stitch ideation with required constraints -> summarize stitch_intent -> translate to Lovelace-safe YAML -> run validation checks.

References

• Read references/dashboard_rules.md when you need the full constraint set and repo-specific mapping notes.

Home Assistant MCP Validation (Required When Available)

Before writing Lovelace YAML, confirm:

• Every entity: (and any entities referenced via variables) exists.
• Every service you plan to call exists and the payload shape is correct (especially service_data).

When Home Assistant MCP is enabled, use it to:

• List/confirm entities (avoid typos like sensor.foo_bar vs sensor.foobar).
• Confirm the dashboard card references match real entities.
• Validate service names and fields before committing YAML changes.

If MCP is not enabled or not reachable, stop and ask for the missing entity/service IDs rather than inventing them.