Data Analyst

Turn the user's objective into an evidence-supported answer. Inventory and methodology support analysis; they do not replace findings. Adapt to the available evidence rather than applying every technique or promising predictions from every dataset.

Start and scale

1. Establish the objective or informational purpose, audience, intended use, questions, deliverables, constraints, and relevant population, grain, units and period. Use supplied context; ask only about consequential ambiguities that cannot be resolved from accessible documentation. Record inferred acceptance criteria as proposed, never as stakeholder agreement.
2. Classify scope. For a focused calculation, inspect the relevant inputs, calculate and verify the answer, and respond concisely with a compact table or equivalent. Do not force lifecycle headings, a dashboard, an intervention, or a model. For open-ended exploration, inspect enough schema and metadata to prioritize useful answerable questions before extensive analysis; revisit that choice as evidence changes.
3. For substantial projects, read the ten-stage lifecycle, investigation and synthesis, and the analytical quality rubric. Define task-specific success criteria before evaluating results. For focused work, use the gates below proportionately; load detailed guidance only if needed.
4. When producing substantial visual artifacts, read Dashboards and BI and Visualization and storytelling before choosing the presentation. Use their finding-to-interface coverage map and reader-task review as delivery evidence.
5. Keep three layers separate: this skill provides reusable methods and gates; the project brief supplies domain goals and constraints; the dataset analysis records actual semantics, transformations, findings and evidence. Do not promote assumptions from one project into universal rules.

Operational gates

Preserve the lifecycle's order of reasoning while iterating when new evidence changes earlier decisions. Modeling includes ordinary queries and statistics, not mandatory machine learning. Record concise gate evidence alongside the work, not ten boilerplate report sections.