| name | stack-selection |
| description | Select a dashboard-engagement stack via the Case A/B/C/D Mermaid decision tree (Portfolio / Per-client / Productized SaaS / Pipes-only) — surfaces the per-viewer-pricing-trap heuristic, recognizes the EdTech LMS connector-gap, returns a populated `stack-decision-record.md`. Invoked by `ravenclaude-core/architect` via inline prior. |
Skill: stack-selection
Invoked by: ravenclaude-core/architect (via inline prior on the core architect's file) — and by any data-platform agent who realizes the engagement Case hasn't been named yet.
When to invoke: any time the user asks "what stack should I use?", a new dashboard engagement is starting, or scope materially shifts mid-engagement. Run this before the database / ELT / dashboard / embed-pattern decisions are made — those decisions hang on the Case outcome.
Output artifact: populated ../../templates/stack-decision-record.md.
The four Cases
Every dashboard engagement fits one of four cases. The skill's job is to name the Case, then ship the default stack for that Case (with documented alt-stack triggers).
Case A — Portfolio dashboard on the consultant's own marketing site
- Shape: Build-as-code, version-controlled, periodic refresh (monthly/quarterly), public-facing, single-tenant (no viewer scope).
- Driving question: How do I show what I can do?
- Default stack: Evidence.dev OSS (MIT-licensed) → static deploy to Vercel / Netlify. Markdown + SQL fenced blocks. $0 hosting.
- Alt-stack triggers: Observable Framework (open-source) if the dashboard needs richer notebook-style data exploration; a custom Next.js + Tremor + Recharts page if brand differentiation matters and Evidence's styling is too generic.
- Don't: Use Evidence Cloud (no free tier; Pro $25/user/mo; Embedded is Enterprise-only). OSS Evidence handles Case A cleanly.
Case B — Per-client deliverable dashboard inside the client's app
- Shape: Multi-tenant or single-tenant (client-scoped), JWT-embedded, 5-50 viewers per client, lives inside the client's existing web app or admin panel.
- Driving question: How do I ship a custom dashboard for this engagement?
- Default stack: Apache Superset OR Metabase OSS, self-hosted on $20-40/mo VPS per client, JWT-secured embed SDK, Postgres RLS for tenant isolation.
- Alt-stack triggers:
- Microsoft-stack client → Power BI Embedded F2 (~$262/mo PAYG, ~$156/mo reserved). App-Owns-Data flow; no per-viewer license; DAX-role RLS in the semantic model. Coordinate with
power-platform/power-bi-engineer.
- Client already on Snowflake / Databricks → data sharing, not a pipeline. Skip the iPaaS entirely.
- Tight budget + read-only → Evidence.dev OSS static deploy if the dashboard doesn't need interactive filtering.
- Don't: Metabase Pro ($575/mo + $12/viewer) — the math breaks at 5-50 viewers. Tableau Embedded (
$420/viewer/yr). Looker ($400/viewer/yr). Sigma ($61k median deployment). All per-viewer-priced; all wrong for SMB consulting.
Case C — Long-bet productized SaaS
- Shape: Multi-tenant, configurable, customer-facing inside the client's own admin panel, 50-500+ viewers, eventually scales.
- Driving question: How do I build a dashboard product I can sell to many?
- Default stack: Cube OSS (Apache 2.0 semantic layer) + custom React UI (Next.js + Tremor + Recharts + shadcn/ui) + Postgres with RLS.
- Alt-stack triggers:
- Cube Cloud Premium ($80/dev/mo, includes embedded dashboards) if running Cube as a service vs. self-hosting.
- Embeddable.com (managed semantic-layer alternative; sales-quoted) if 5+ paying tenants and operating the platform layer is the bottleneck.
- Migrate hot-path queries to Apache Druid / Pinot / ClickHouse Cloud at very high scale (Stripe runs Pinot at p99 70ms / 10k+ QPS — the "big version").
- Don't: Skip the semantic layer. Customer-facing dashboards shipping raw SQL queries will not scale.
Case D — Client already has a BI tool; scope is data pipes only
- Shape: Client has Power BI / Tableau Online / Looker Studio / Metabase already deployed. Engagement is ELT + database + maybe modeling — no dashboard build.
- Driving question: How do I get the right data into the client's existing tool?
- Default stack: Airbyte (Cloud Standard or self-hosted) + Supabase Pro or Postgres on the client's cloud + dbt Core for modeling. No dashboard framework selection.
- Alt-stack triggers: Fivetran free tier if MAR <500K AND client takes over post-engagement.
- Don't: Try to sell a dashboard build the client doesn't need. Stay in the data-pipes lane.
Decision Tree: Dashboard engagement — which Case?
When this applies: any new dashboard engagement OR mid-engagement scope shift that might change the Case. Traverse this tree top-to-bottom before recommending a stack — do NOT pattern-match on keywords in the user's situation description.
Last verified: 2026-05-22 against current pricing pages + practitioner sources.
flowchart TD
START[New dashboard engagement] --> Q1{Portfolio for the consultant's own marketing site?}
Q1 -->|YES| CASE_A["Case A — Portfolio<br/>Evidence.dev OSS<br/>Static deploy"]
Q1 -->|NO| Q2{Client already has a BI tool — needs only data pipes?}
Q2 -->|YES| CASE_D["Case D — Pipes only<br/>Airbyte + Postgres + dbt<br/>No dashboard framework"]
Q2 -->|NO| Q3{One-off custom dashboard inside the client's app?}
Q3 -->|YES| Q3a{M365-stack client?}
Q3a -->|YES| CASE_B_PBI["Case B — Microsoft variant<br/>Power BI Embedded F2 ~$262/mo<br/>App-Owns-Data + DAX-role RLS"]
Q3a -->|NO| CASE_B_OSS["Case B — OSS variant<br/>Superset or Metabase OSS<br/>VPS $20-40/mo + Postgres RLS + JWT embed"]
Q3 -->|NO| Q4{Multi-tenant product to sell to many customers?}
Q4 -->|YES| CASE_C["Case C — Productized SaaS<br/>Cube OSS + custom React<br/>Postgres RLS"]
Q4 -->|NO| ESCAPE["None of A/B/C/D fit<br/>Escalate to ravenclaude-core/architect"]
Rationale per leaf:
- CASE_A — single-tenant + version-controlled + public-facing means no embed-auth complexity. Evidence.dev OSS handles the SQL-fenced-block authoring pattern cleanly. $0 hosting via Vercel / Netlify static deploy.
- CASE_D — when the client owns the BI tool, the dashboard build is out of scope. Stay in the pipes lane. Don't try to sell a dashboard the client doesn't need.
- CASE_B_PBI — Microsoft-stack means Entra-ID-based RLS via DAX roles is the path of least resistance. F2 capacity is flat ~$262/mo PAYG (no per-viewer license). Coordinate with
power-platform/power-bi-engineer on semantic-model + DAX-role design.
- CASE_B_OSS — non-Microsoft client + 5-50 viewers means per-viewer pricing kills the math. Superset / Metabase OSS with Postgres RLS + JWT embed is the consulting-friendly default at $20-40/mo per-client VPS.
- CASE_C — multi-tenant productized SaaS needs a semantic layer (Cube) to prevent customer-facing dashboards from shipping raw SQL. Custom React (Next.js + Tremor + Recharts + shadcn/ui) gives brand differentiation and viewer-scale economics.
- ESCAPE — if the engagement genuinely doesn't fit A/B/C/D, the architect handles a custom plan. Examples: regulator-facing dashboard with audit-trail requirements; embedded-mobile-app dashboard; offline-first dashboard for field operations.
Tradeoffs summary:
| Case | Stack | Per-month cost | Multi-tenant? | Viewer ceiling |
|---|
| A | Evidence.dev OSS + static deploy | $0 | No (single-tenant) | n/a — public |
| B (OSS) | Superset/Metabase OSS + VPS + Postgres RLS | $20-40/client | Yes | 5-50/client |
| B (PBI) | Power BI Embedded F2 + Fabric | $262 PAYG / $156 reserved | Yes (Entra-ID RLS) | Unlimited |
| C | Cube OSS + custom React + Postgres RLS | $0 OSS + infra (~$50-200/mo at scale) | Yes | 50-500+ |
| D | Airbyte + Postgres + dbt | $20-40 (Airbyte Cloud) or self-host | Client-owned | n/a — client's BI tool |
Failure modes to avoid:
- Recommending Metabase Pro / Tableau Embedded / Sigma / Looker for Case B without flagging the per-viewer math (5-50 viewers × $400+/yr each kills SMB consulting margin)
- Putting Case C on a raw-Postgres stack without a semantic layer (Cube). Customer-facing dashboards shipping raw SQL won't scale.
- Recommending Case A's Evidence.dev for a multi-tenant Case B engagement (no embed-auth model in OSS Evidence)
- Trying to retrofit Case C onto a Case B engagement just because the consultant might sell it to a second client someday — Case C earns its complexity only when there ARE multiple paying tenants
The per-viewer-pricing-trap heuristic
When the user names a BI tool, run this check before recommending it for Case B or C:
| Tool | Viewer-cost model | When math breaks |
|---|
| Looker (enterprise) | ~$400/viewer/yr | 5+ viewers per client × 4-6 clients |
| Tableau Embedded | ~$420/viewer/yr | Same |
| Sigma | Median $61k deployment, embedded 2-3× | Any SMB consulting profile |
| Metabase Pro Interactive Embedding | $575/mo + $12/viewer | 15+ viewers |
| Power BI Embedded F2 | $262/mo PAYG flat capacity (no per-viewer in App-Owns-Data) | Doesn't apply — flat |
| Apache Superset OSS | Free | Doesn't apply |
| Metabase OSS (static embed only) | Free | Doesn't apply |
| Cube OSS | Free; Cloud per-developer | Doesn't apply |
| Evidence.dev OSS | Free | Doesn't apply |
The hook (../../hooks/flag-data-platform-smells.sh) enforces this on stack-decision-record.md templates — references to Looker / Tableau Embedded / Sigma / Metabase Pro trigger a stderr warning.
EdTech LMS connector-gap recognition
When the engagement is EdTech and a connector is needed for Canvas / Moodle / Schoology / Blackboard / D2L: this is the canonical custom-Airbyte-connector use case. Native ELT vendor coverage is thin. Surface this finding in the decision record + route the partner-success motion (renewals, QBR, health scoring above the data layer) to edtech-partner-success agents. See ../../knowledge/edtech-lms-connector-gap.md.
Inputs the skill needs from the requester
Before recommending, confirm:
- Engagement target user: consultant's own site / single-client deliverable / multi-tenant product / data-pipes-only?
- Client tech stack: Microsoft (Dynamics, Fabric, Power BI), Salesforce, generic SaaS, AWS-native, GCP-native, Snowflake/Databricks already?
- Compliance: HIPAA, SOC 2, GDPR, state-privacy (NY Ed Law 2-d, IL SOPPA, CA SOPIPA)?
- Source systems: what data is in play? (QBO, Stripe, Salesforce, HubSpot, GA4, Shopify, LMS, HRIS, other?)
- Viewer count: 1-5 / 5-50 / 50-500 / 500+?
- Engagement length: one-off / multi-month / ongoing retainer?
- Post-engagement ownership: client takes over the infra, or Matt hosts indefinitely?
These seven inputs drive the Case match + the alt-stack triggers.
Anti-patterns this skill flags
- Picking a framework before naming the Case
- Recommending a per-viewer-priced BI tool for Case B/C without surfacing the cost math
- Defaulting to Snowflake / Databricks for engagements <$25K ACV
- Pulling a new ELT pipeline when the client is already on a lakehouse (data sharing is the answer)
- Skipping the semantic layer for Case C (customer-facing dashboards shipping raw SQL won't scale)
- Recommending Evidence Cloud for Case A (no free tier; Embedded is Enterprise-only — OSS handles Case A)
- Recommending Metabase Pro Interactive Embedding without explaining the $144/viewer/yr + $575/mo base math
- Failing to flag the Fivetran 2026 deletes-count-as-MAR change when proposing Fivetran on change-heavy sources
Output: stack-decision-record.md
The skill ships a populated ../../templates/stack-decision-record.md with:
- Case match (A/B/C/D) + rationale
- DB / ELT / dashboard / embed picks per the Case defaults (with alt-stack overrides documented)
- Pricing claims with retrieval dates
- Compliance constraints flagged
- Cross-plugin handoff routes (
power-platform, edtech-partner-success, web-design)
References
