Lab: The refund story
Turn the Module 2 memo into the full readout package: three chart specs, an answer-first one-pager, the Q&A prep, and the dashboard spec for what comes after.
You have the findings; this lab produces the artifacts that make them move: chart specs, the one-pager, Q&A prep, and the follow-up dashboard spec. Grade yourself against the checklists — they're the rubric a senior analyst would use on your work.
- 1Spec the three evidence charts with the chart-spec prompt: (1) the localization bar (categories, Gear highlighted), (2) the bend line (Gear web vs. store, April annotated), (3) the mechanism stack (reason mix by month). Titles must be claims. Build them in your spreadsheet from the lab extracts — specs are cheap; the chart you actually build finds the label collision the spec missed.
- 2Write the one-pager: answer sentence → three beats (one per chart) → the ask (fragile-flag audit + Gear packaging exception, owner, date) → confidence line (the pending audit, stated plainly) → appendix pointers (alternatives killed, numbers log). Hard limit: one page. If it doesn't fit, you're presenting the appendix.
- 3Run the three-persona Q&A simulation. For every question the AI raises that your package can't answer, either add the answer to the appendix or add it — honestly — to the confidence section. 'I don't know yet, and here's when I will' is an allowed answer; being surprised is not.
- 4Draft the re-framing pair: the CFO version and the COO version of your answer sentence. Same finding, different first clause. Notice how little changes — and how much.
- 5Spec the follow-up dashboard (5 tiles max): refund % by category-channel (the cell that caught this), damage-reason share, and whatever else earns a tile. For each: the metric definition it uses and who owns that definition. Five governed tiles beat twenty vibes.
Hand your one-pager to someone who knows nothing about the investigation for ten seconds, take it back, and ask what they retained. If the answer isn't approximately 'broken gear, packaging thing in April, they want an audit' — the page isn't done. This test is brutal, free, and never wrong.
You get a deliberately awful five-chart deck (truncated axes, dual y-axes, a six-slice pie, a title that's a caption, an unmarked definition change mid-series) to critique and rebuild, plus a one-pager written analysis-first to resequence answer-first. Editing bad work is how the standards become reflexes.