This page hosts the living version of The Defensible Decision's Appendix B prompt library. The eleven prompts below are paired with Chapters 8 through 14 and each maps to a specific Copilot surface. Each prompt has its own page with the full CSAR-loop walkthrough (Crystallize / Scope / Assemble / Refine) and the ship-decision criterion.
Every prompt is available as both Markdown and Word (.docx) — easier than typing them from the print edition. Markdown pastes cleanly into Copilot; Word is for review workflows, team sharing, and the “actually just open it in Word” cases. Each character-anchored entry links to the published companion dataset so you can pull the same data the anchor practitioner used and reproduce the prompt end-to-end.
The point of the library is not to copy and paste the prompts verbatim. The prompts are tied to specific semantic models, audiences, and decisions; pasting them into a different context will produce different output. The point is the shape of each prompt, the placement of each Refine correction in the loop, and the vocabulary the analyst uses to name the rubric. Read the library as a pattern catalog, not a script.
How to use this library
Each prompt is pinned to a specific Copilot surface and a specific verification date. Copilot ships surface changes on a roughly monthly cadence; a prompt that worked in June 2026 may behave differently in October 2026 because the surface itself has changed, not because the prompt is wrong.
Treat the verification date as the prompt's freshness signal. When you encounter unexpected behavior, check the date first; if it's more than three months old, the surface may have shifted. Report drift via the errata page so the next revision can incorporate the fix.
The eleven canonical prompts
1. Build an Executive Summary
The CSAR-loop entry point for prose-heavy summaries built from a semantic model. Source: Sara worked example (EMEA Q3 miss).
2. Generate a Comparison Report
Plan-vs-actual reports against a strict finance measure. Source: Diego worked example (campaign attribution).
3. Surface Drivers for a Metric
The pattern for the Key Influencers AI visual. Source: Asha worked example (escalation drivers). Includes the confound-check move that distinguishes ranking-acceptance from operational lever identification.
4. Detect Anomalies in a Time Series
The pattern for the Anomaly Detection AI visual. Source: Asha's volume-line-chart pass. Includes the business-context check that distinguishes real signals from publicized-rollout false positives.
5. Explore a Metric with an AI-Assisted Tree
The pattern for the Decomposition Tree AI visual. Source: Asha's escalation-page pass. Includes the filter-clear re-walk that catches paths dependent on the analyst's slicer state rather than the audience's view.
6. Summarize a Report Page in Prose
The pattern for the Smart Narrative AI visual. Source: Asha's Smart Narrative pass. Includes the BRD-takeaway alignment check that catches drift — the highest-risk failure mode because it reads correct.
7. Ask a Question of the Data
The natural-language Q&A surface and the demonstration of why Prep for AI matters. Source: Mei worked example (revenue-cycle model, before and after the three configurable Copilot surfaces). The same prompt produces different answers depending on the model's prep state.
8. Build a Dashboard from a Brief
The CSAR-loop entry point for full-page dashboards anchored on a primary KPI. Source: Rohan worked example (executive dashboard). The discipline is in resisting Copilot's tendency to add and in cutting until the page reads cleanly at thumbnail size.
9. Recognise a Vibe-Charting Session
Diagnostic entry: the prompt-roulette failure mode side-by-side with the criteria-anchored disciplined alternative. Source: Yuki worked example (1:47 a.m. dashboard polish, five-of-five diagnostics).
10. Translate Question to Defensible Artifact
The cross-surface pattern that converts causal questions into what questions before the artifact gets built. Applies regardless of which Copilot surface the question lands on — the most cross-cutting prompt in the library.
11. Four Responsible-AI Moves Audit
The full audit pattern that runs after the CSAR loop, before any high-stakes AI-generated artifact ships. Source: Anand worked example (headcount-allocation dashboard). Provenance, fairness, decision boundaries, and disclosure applied in sequence so each move's output changes the next.
Patterns across the library
Eleven prompts across eight scenarios: seven character-anchored worked examples (Sara, Diego, Asha's four AI-visuals pass counted once, Mei, Rohan, Yuki, Anand) plus the cross-surface pattern. Four patterns recur across all of them:
- Crystallize always names the measure, the grain, and the takeaway. When any of the three is missing, the Refine column gets longer. On the AI-visual surfaces (prompts 3–6), the Crystallize move condenses to the field selection and the initial filter state — no conversation, but the same three commitments have to be made explicitly before the visual returns anything.
- Scope is the cheapest phase to skip and the most expensive omission. Seven of the eleven entries include an explicit conversational restatement turn; the four AI-visuals entries do not (the surface produces output directly), and the analyst has to stand in for the Scope move by writing the BRD takeaway or the field-set audit down on paper before opening the visual.
- Refine cites a Part I principle by name. A correction without a principle attached is a vibe-charting correction; it tells the analyst nothing about how to prevent the same defect on the next artifact. For entries where a chapter-specific per-visual move exists (Chapter 10's four AI-visuals moves, Chapter 13's five diagnostics, Chapter 14's four responsible-AI moves), the Refine column cites both the Part I principle and the chapter-specific move.
- Every character-anchored entry runs on a published dataset. Each entry links to the companion dataset so you can pull the same data the anchor practitioner used and reproduce the prompt end-to-end. The cross-surface pattern (prompt 10) is deliberately dataset-agnostic; it applies wherever a causal why needs converting to a data-answerable what.
Suggestions and errata
The library will continue to grow as new Copilot surfaces ship and as new worked examples surface in revisions of the book. If a prompt no longer behaves as documented against the current Copilot surface, or if you want to suggest a new pattern for a surface not yet covered, report it via the errata page.