Chapter 7 — Practice-on-your-own solutions

How to read this worked example. This is a reflective problem, not a graded one. The example below is one plausible walk — a mid-career analyst who works mostly in Power BI and Excel and admits that context (Chapter 2) is the discipline they skip most often. Your own blind spot will be different, and your predictions for Part II should trace back to your weakest principle. What matters is the honesty of the self-audit, not the match to this specific answer.

Worked example — weakest principle: Chapter 2 (context). The analyst reports: «I skip the BRD. I open the dataset, look at the numbers, pick a chart family that feels right, and iterate from there. I know I should elicit the decision first and the comparison second, but the request usually arrives as ‘can I get a chart of X’ and I go straight to the chart.»

Predicted Part II failure modes (one sentence per chapter):

• Ch 8 (author-not-typist): I will accept Copilot’s first-draft chart because I have no BRD to test it against — the chart looks fine on its face and I have no way to name what is missing.
• Ch 9 (prompt as spec): My prompts will underspecify the audience and the decision, so Copilot will fill the gap with its defaults and I will end up directing whatever it happens to produce rather than what the audience needs.
• Ch 10 (three prompts): The plan/setup/refine prompts will not have the load-bearing constraints (audience, decision, comparison) in them, so the refinement loop will polish the wrong chart.
• Ch 11 (audit the AI): My audit will focus on visual defects (colour, sort order, labels) and miss the structural defect — that the chart is answering the wrong question.
• Ch 12 (defensibility): When Rohan-style questioning happens in a meeting, I will not have a BRD to point to and my defence will collapse into «that’s what the data showed.»
• Ch 13 (vibe charting): My reroll loop will not have a rubric because I never authored a BRD to derive one from; I will vibe-chart the rebuild.

What the exercise is for. The prediction is the audit’s payoff. When you get to each Part II chapter, you can check whether the failure mode you predicted actually showed up. If it did — and it usually does — the weakest-principle audit was calibrated correctly. If a different failure mode showed up, the audit found a second blind spot the first pass missed. Either way, the return trip in Chapter 13 has data to work with.

The audit format. For each of the three charts, record five things: (1) the chart’s decision context in one line, (2) the stakes (low / medium / high), (3) what the AI produced on first draft, (4) what you did in response, (5) the role — author, director, or bystander.

Worked example — three recent charts from a mid-career analyst.

1. Weekly ops summary for the team lead. Stakes: low (recurring, low-visibility). AI produced a stacked bar of tickets-by-status-by-team. I changed the title from «Ticket status» to «Weekly ticket volume, w/e 9 Aug» and shipped it. Role: bystander.

2. Quarterly review deck slide for the VP. Stakes: high (visibility to leadership, informs next-quarter planning). AI produced a clustered column of revenue-by-quarter-by-product-family. I asked for a plan-variance chart instead, sorted ascending, with the underperforming product families highlighted. I re-annotated the callouts against the BRD. Role: director.

3. One-page executive summary for the CFO on Q3 spend. Stakes: high (single-page, single-decision, one-shot artifact). AI produced a table and a treemap. I rejected both, wrote the Big Idea in one sentence, chose the chart family myself (a slope graph of budgeted vs actual per cost centre), and iterated the chart from the Big Idea forward. Role: author.

The pattern. The role scaled with the stakes: bystander on low-stakes recurring work, director on medium-visibility set-pieces, author on the highest-stakes single-shot artifacts. That is the correct calibration — the discipline should be heaviest where the cost of shipping the wrong chart is highest.

The failure mode to watch for. The dangerous pattern is bystander on high-stakes charts. If any of the three high-stakes rows had been bystander, the next chart cycle needs the same intervention Chapter 12 walks through: author the BRD before you open the AI surface, treat the AI’s first draft as one of several candidates rather than the starting point, and expect to spend most of the chart’s build time on the last twenty percent of decisions — the ones that make the chart defensible instead of merely presentable.

If the pattern is bystander across the board. The self-audit has surfaced the same finding Yuki’s session (Chapter 13) surfaces from a different angle: vibe-charting has become the default. The next chart cycle needs a rubric written before the first prompt, a verification prompt included in the loop, and a hard stop where you compare the shipped chart against the rubric before hitting send. Chapter 13’s diagnostics and Chapter 12’s defensibility test give you the vocabulary to run that stop.