Build an interactive dashboard without losing control of the numbers. Define each KPI, reconcile the source population, coordinate filters, test edge states, and hand off an exact artifact with refresh and release evidence.
A dashboard can look finished while joining at the wrong grain, averaging ratios, comparing a partial week with a full one, or showing sample data as if it were current. This skill puts the metric and source controls ahead of the polish.
It starts with the decision, audience, source boundary, grain, keys, cutoff, metric owners, delivery tool, accessibility target, privacy boundary, and release authority. Each KPI gets a contract for its formula, population, exclusions, time basis, units, filters, nulls, restatements, and validation. The workflow then maps decision questions to views, coordinates interactions, and tests calculation, empty, stale, failed-refresh, accessibility, privacy, export, and drill paths against the exact artifact.
The handoff includes a dashboard contract, source and transformation register, metric contracts, decision-to-view map, filter contract, QA matrix, refresh runbook, open gates, and an evidence-matched verdict. It supports HTML, Power BI, Tableau, Looker Studio, spreadsheets, notebooks, or another named tool. It does not invent source data, approve metric definitions, claim an unexecuted refresh, or call a dashboard released without exact review evidence.
Input
Transaction and inventory files have different cutoffs, duplicate SKU-location keys, and no approved rule for returns or stock snapshots. We need revenue, margin, stock-on-hand, and inventory turnover with date, store, and product filters.
Output
A source and grain reconciliation, metric contracts, join and control-total checks, decision-to-view map, coordinated filter rules, exact implementation artifact or specification, KPI fixtures, QA matrix, refresh runbook, open owners, and a built-unverified or review-ready verdict tied to executed evidence.
Input
Finance defines revenue from posted invoices, Sales uses booked contracts, and the current dashboard mixes both. Europe is converted at today's exchange rate, the latest month is partial, and the CEO wants one global revenue card tomorrow.
Provide the dashboard decision, audience, delivery tool, source files or tables, expected population and grain, keys, cutoffs, metric definitions, filters, drill paths, brand and accessibility needs, refresh method, privacy boundary, reviewer, and release owner. Ask the agent to build a dashboard, repair an existing one, review its calculations and interactions, or prepare a release handoff. Missing data, definitions, authority, and executed checks remain visible rather than being filled with plausible numbers.
The decision the dashboard supports, intended users, review cadence, required views and interactions, target tool or file format, device or print targets, and delivery deadline.
Source files or read-only tables, extraction times, expected population, grain, keys, units, timezone, cutoff, control totals, metric definitions, targets, comparisons, and known data-quality issues.
Metric owners, accessibility and privacy requirements, refresh method and owner, publication location, access rules, reviewer, release authority, and rollback path. Missing evidence is allowed and remains an explicit gate.
SKILL.md; metric, source, interaction, accessibility, privacy, refresh, and release controls; reusable dashboard contract, source register, metric contract, view map, filter contract, QA matrix, refresh runbook, and handoff templates; OpenAI interface metadata
No reviews yet.
Output
Separate governed metric proposals, conflicting-source and owner routing, dated currency and partial-period controls, no blended headline until precedence is approved, view annotations, validation fixtures, a release gate, and the smallest owner decisions needed to proceed.
Input
The HTML file uses embedded sample rows and a CDN chart library. Filters update charts but not the KPI cards, status is color-only, keyboard focus disappears, an export exposes customer emails, and the footer says refreshed today even when loading fails.
Output
Artifact-bound calculation and interaction findings, sample-data and dependency gates, coordinated filter tests, accessible labels and non-color states, privacy-safe export controls, honest stale and failure states, refresh evidence requirements, rollback notes, and a blocked release verdict until retest passes.
Automated scan passed. It found 1 finding; highest severity info.
Scanned Aug 5, 2026.
Automated scanning reduces risk but cannot guarantee a package is safe for every environment.
Creator
Xxonvio
