powerbi-dashboard-review

Power BI dashboard review

Find what is wrong with a report before a director does. Arithmetically correct numbers that still tell a false story are the whole reason this skill exists — docs/reference/PLAYBOOK.md 1.1. The reviewer is read-only: it produces a findings table and a change plan, and never edits or deploys.

Procedure

1. Write one shared brief before anything else: what the report is, who reads it and who has seen it, a digest of what each visual really queries — read out of the PBIR JSON via get_report_definition, not out of visual titles — what is on screen today as plain text, the facts already established from the model with every hypothesis marked as one, the hard constraints (model read-only, no refresh, theme locked or open, "the audience will not touch a slicer"), and what has already been fixed. Give every lens in step 4 this same brief, so disagreement is about meaning, not data.
2. capture_pages(target, out_dir, ...) for every page in its default state first — nothing clicked, because for an executive report the default state is the product. Then repeat for each slicer pick and the reset button. Do not skip straight to slicer states; the default-state screenshots are what most readers actually see.
3. Pull ground truth with one batched run_dax(query, ...) — latest traded day, week-to-date, year-to-date, last year same weekday, budget, per channel. One query, not one per number (docs/reference/PLAYBOOK.md 2.6, 2.8 — capacity etiquette applies to a review too).
4. Run five lenses over the same brief, screenshots and ground truth, each returning a one-paragraph verdict plus findings in this record: id · title · page · severity · lenses · what_exec_sees · why_it_matters · proposed_change · effort · evidence · theme. what_exec_sees is the discipline: it forces a quote of the screen, not the JSON. The questions each seat asks in its first ten seconds: