BI Dashboards skill for Claude Code

A dashboard is done when each number on it traces to one written, owned metric definition, the layout answers a stated question in the order people ask it, filters and permissions behave predictably, it loads in a few seconds, and someone is accountable for it. Charts come last.

Workflow

- [ ] 1. Write the dashboard brief: audience, decisions, questions, cadence
- [ ] 2. Define each metric in the metric register
- [ ] 3. Build or reuse a governed dataset / semantic model
- [ ] 4. Lay out: headline KPIs, trends, breakdowns, detail
- [ ] 5. Add filters, drill paths, and row-level security
- [ ] 6. Set refresh/caching; test load time
- [ ] 7. Validate numbers against a trusted source; certify; assign owner

1. Brief

**Dashboard:** Weekly revenue health      **Owner:** finance-analytics (Dana)
**Audience:** VP Sales, regional leads    **Cadence:** Monday review, data as of Sunday 23:59 UTC
**Decisions:** where to shift sales capacity; which regions need pricing review
**Questions:** Are we on plan this quarter? Which regions/segments are behind? Is it volume or price?
**Out of scope:** pipeline forecasting (separate dashboard)

If the requester cannot name a decision the dashboard supports, propose a one-off analysis instead of a permanent dashboard.

2. Metric register

Metric	Definition	Grain / source	Filters baked in	Owner
Net revenue	sum(net_amount) of invoices, excluding tax, after refunds, in USD at invoice-date rate	fact_invoice_line	status in (paid, partially_refunded)	Finance
Active customers	distinct customers with >= 1 paid invoice in the period	fact_invoice	-	Finance
ARPA	net revenue / active customers, same period	derived	-	Finance

Rules:

• Ratios are computed as sum(numerator) / sum(denominator) at the displayed level, never as an average of row-level ratios.
• State the time attribution (invoice date vs payment date vs order date) and time zone.
• Define once in the semantic layer; dashboards reference the definition, not copy SQL.
• When two teams use the same name for different things, rename one (“bookings” vs “recognized revenue”) instead of tolerating ambiguity.

3. Semantic layer per tool

Tool	Where definitions live	Notes
Metabase	Models (curated queries with column metadata) and Metrics; questions built on models	verify/certify models; keep raw tables hidden from casual users
Superset	Datasets (physical or virtual) with metrics and calculated columns	define metrics on the dataset, not in each chart; cache per dataset
Power BI	semantic model: star schema in Power Query + DAX measures	one shared semantic model, thin report files on top
Looker Studio	data source with calculated fields; ideally a prepared BigQuery view	blended sources are fragile and slow - pre-join in the warehouse
Looker	LookML views and explores	measures defined in LookML, versioned in git

Push heavy logic into the warehouse (dbt models or views) and keep BI-layer logic to measures and formatting. Model the source as a star schema (see dimensional-modeling).

Power BI measure examples:

Net Revenue := SUM ( fact_invoice_line[net_amount] )
Active Customers := DISTINCTCOUNT ( fact_invoice[customer_key] )
ARPA := DIVIDE ( [Net Revenue], [Active Customers] )
Net Revenue PY := CALCULATE ( [Net Revenue], SAMEPERIODLASTYEAR ( dim_date[date] ) )

Use measures, not calculated columns, for aggregations; mark dim_date as the date table; relationships single-direction from dimension to fact unless there is a proven need.

4. Layout

Read order is top-left to bottom-right; place by question priority:

1. Headline row: 3-5 KPI tiles - value, comparison (vs plan / prior period), and sparkline. Label the period explicitly (“QTD through Oct 19”).
2. Trend: the primary metric over time with target line.
3. Breakdowns: by the dimensions in the brief (region, segment), sorted by value or by gap to plan, not alphabetically.
4. Detail: a table for drill-down, paginated, with export.

Rules: one question per chart; consistent colors for the same category across charts; no more than ~10-12 charts per page (split into tabs/pages beyond that); show “data as of” timestamp; annotate known incidents (tracking outage, price change). Chart choice and encoding follow data-visualization.

5. Filters, drill, security

• Global filters (date range, region) at the top; show which charts each filter affects. Default the date filter to the period in the brief.
• Cross-filtering and drill-through (Power BI drillthrough pages, Superset drill to detail, Metabase click behavior) beat duplicating charts per segment.
• Row-level security in the tool (Power BI RLS roles with DAX filters, Superset RLS filters, Metabase data sandboxes, Looker access filters) or in the warehouse via authorized views. Test with “view as” for at least two roles.
• Never rely on hiding a filter or a page as a security control.

6. Performance and refresh

Symptom	Fix
Every chart runs a heavy query on open	pre-aggregate in the warehouse (daily summary table), cache results
Looker Studio slow on blends	replace blends with a warehouse view; use extracts for small data
Power BI DirectQuery slow	import mode with scheduled refresh, or aggregations tables
Dashboard shows mixed freshness	align refresh schedules to the upstream pipeline completion

Target: under 5 s to first meaningful paint for common filters. Schedule refresh after the pipeline that feeds it, not on a fixed clock that races it.

7. Governance

• Certify (verified/endorsed badge) only dashboards whose metrics come from the register.
• Each dashboard has an owner, a description with the brief, and a “last validated” date.
• Organize by domain folders with a clear “certified” vs “sandbox” split.
• Retire content: list dashboards with no views in 90 days (usage analytics in each tool), notify owners, archive after confirmation rather than deleting.
• Version where possible: LookML and Power BI projects (PBIP) in git, Superset and Metabase exports/serialization for promotion between environments.

Gotchas

• Distinct counts do not add up across rows; a “total” row summing distinct customers per region overstates the company total.
• Timezone mismatch between warehouse (UTC) and tool settings shifts daily numbers near midnight.
• Default aggregation guesses (Looker Studio, Superset) sum IDs or average averages; set aggregation explicitly on every field.
• Partial current period plotted next to full periods looks like a drop; mark or exclude it.
• Filters on a dimension with a Type 2 history join can restate the past; decide whether reports show attributes as-of-event or current.

Verification

• Reconcile each headline KPI against the trusted source (finance close, billing system) for two closed periods; record the difference and explanation.
• Test every filter combination in the brief and each RLS role via “view as”.
• Load time measured for default and heaviest filter settings.
• Metric register entries linked from the dashboard description.