Imported from CatCorner22/Claude_Skills_2 (plugins/decision-science-skills/skills/minority-report/SKILL.md). Install upstream with npx skills add CatCorner22/Claude_Skills_2 --skill minority-report. Copyright stays with the author.

Minority Report (the precog cell)

Three precogs, one tank — and the system's fatal flaw was never the visions. It was the filing cabinet: when the three reports disagreed, the majority went to the ball and the minority report went in a drawer. This skill is that machine rebuilt with the drawer welded open. It generates multiple realistic futures for whatever you are deciding, shows you which variable each future turns on, and treats the dissenting scenario as the most valuable page in the file — because the futures you rehearse are the ones that cannot ambush you.

When to use

• A decision needs its future outcomes explored: a project direction, an offer, a launch, a career move, an architecture choice, a case strategy.
• The user asks "what happens if X changes" — a variable (funding, timeline, a ruling, a hire, demand) needs turning to see which outcomes flip.
• A plan rests on one confident picture of the future and nobody has written down what else could happen.
• Not for: playing a thinking adversary move-by-move → decision-science-skills:tabletop-wargaming; imagining the causes of one failure before commitment → decision-science-skills:pre-mortem (a scenario cell holds futures where things go RIGHT differently, too); base-rate estimates for a single quantity → decision-science-skills:reference-class-forecasting (it feeds this skill's probability discipline); statistical extrapolation from history → machine-learning-skills:time-series-forecasting; deciding mid-flight whether a committed plan still deserves its timeline → decision-science-skills:the-challenger (this skill runs BEFORE commitment and re-arms it with tripwires).

Do it

Framing, variable sorting, a domain-neutral worked scenario construction, the minority-report rule, and the one-variable turn are in references/scenario-cell-method.md.

1. Frame the decision, not the future. Write the decision this cell serves, the time horizon, and what would count as a good outcome. Scenario work without a decision attached produces interesting fiction; the decision is what makes it precognition.
2. Sort the driving variables. List everything that materially shapes the outcome, then split: NEAR-CERTAINTIES (trends you can bank on across the horizon) versus CRITICAL UNCERTAINTIES (variables that could genuinely go multiple ways and matter most). The uncertainties become the scenario axes; the certainties appear in every scenario. If systems structure is doing the driving, draft the loops first (the archived decision-science-skills:systems-thinking).
3. Build 3–5 structurally different futures. Not best/expected/worst — that is one future at three volumes. Take the two or three critical uncertainties and build scenarios where they resolve DIFFERENTLY: each future internally coherent, named memorably (names are handles for meetings), narrated concretely (what a Tuesday looks like inside it), with its winners, losers, and second-order effects (the LLM drafts; the human corrects — the correcting is where the thinking happens).
4. File the minority report — always. One scenario must be the dissenter: the future that disagrees with where the room is already leaning, argued at FULL strength, not straw-manned into a token risk paragraph. It gets equal narrative quality, equal indicator coverage, and it is never deleted to make the deck cleaner. The namesake story is about exactly this suppression; the cell exists to make it impossible.
5. Turn one variable. The sensitivity pass, and the user's founding ask: put the scenarios in a working plausibility order first (step 7 disciplines it; a rough order is enough to see movement), then hold everything else fixed, change ONE variable (the deadline slips a quarter; the ruling goes the other way; the key person leaves; demand halves), and record whether the favored CHOICE changes. Rank the options, not the scenario labels: turning a variable that is itself a scenario axis relocates you from one scenario to another by construction, and a flip that cannot come out "no" carries no information. Variables whose turn reorders the choices deserve monitoring money; variables that flip nothing on their own drop to low-priority monitoring — not off the list. One-at-a-time turning is blind to interactions by construction: two variables that each change nothing alone can flip the ranking when they move together, which is the ordinary case when both bear on the same constraint (a supply delay and a demand spike are individually absorbable and jointly fatal). Before retiring a variable, turn it jointly with each of the others that touch the same bottleneck; if the cell has more than a handful of variables, turn the plausible pairs rather than all combinations. Then list the ROBUST MOVES — actions that pay off in every scenario, the minority report included. These are the cell's highest-value product: decisions you can make now without resolving any uncertainty (in the reference's worked case, negotiating data export wins whether you build, keep, or get squeezed).
6. Run the reflexivity check. Acting on this report changes the futures it reports — announce a downturn plan loudly enough and you can cause the downturn; prepare visibly for a risk and you may prevent it (which then looks like the forecast was wrong). For each scenario ask: does the action this report recommends make this future MORE or LESS likely? Predictions that alter their own outcome aren't failures — they are the point. That is Anderton's problem, and it is a real, named property of social prediction (Merton's self-fulfilling and self-defeating prophecies), not a paradox to fear.
7. Attach probabilities honestly or not at all. If a reference class exists, anchor each scenario's likelihood to base rates (decision-science-skills:reference-class-forecasting) and give bands, not point percentages. If no reference class exists, RANK the scenarios by plausibility and say plainly that the ranking is judgment. A confident fake percentage is worse than an honest ordering.
8. Arm the tripwires and log the decision. Each scenario gets 2–3 observable EARLY INDICATORS — things the world would show soon if that future is arriving — plus who watches each one and the threshold that fires it (feed safety-and-reliability-skills:break-glass-playbooks where a fired tripwire needs pre-authored moves). Close with the decision log: what was chosen, which scenario it bets on, which robust moves were taken, what evidence would revisit it, and the next review trigger (decision-science-skills:the-challenger takes it from there).

Why / learn

The real-world anchor is documented practice, not fiction: scenario planning as developed at Royal Dutch Shell (Pierre Wack's planning group; Peter Schwartz's The Art of the Long View) — famously credited with preparing Shell for the 1970s oil shock — replaced single-point forecasts with a small set of structurally different, internally coherent futures, because the value of scenarios is not predicting which one arrives but making the organization UNSURPRISABLE across the set. The method's core insight survives every domain transfer: separating what is effectively certain from what is critically uncertain forces the discovery that most planning conflates the two. Sensitivity analysis supplies the second blade — one-variable-at-a-time turning is how engineers find which inputs actually govern a system — and reflexivity supplies the third: Robert Merton named the self-fulfilling prophecy, and every social forecast lives with it (the PKD story dramatizes it perfectly: Anderton's own foreknowledge changes what he does). The fiction contributes the honesty rule. In Dick's story, the three precogs' reports disagreed, and the system filed the majority and buried the dissent — the plot IS the discovery that the buried minority report was load-bearing. So this cell's non-negotiable is structural: the dissenting scenario is always filed, at full strength, by name. And the fiction contributes one moral, stated once: precrime convicted people for futures that hadn't happened. A scenario cell informs choices; it never convicts anyone — not a person, not a project — on an unhappened future. Forecast to choose, never to punish the prediction. Humans adjudicate; the cell drafts.

Common mistakes

• Best/expected/worst → one future at three volumes; build futures that differ in STRUCTURE (different uncertainties resolving differently), not amplitude.
• Suppressing or straw-manning the dissenting scenario to keep the deck clean → the namesake failure; the minority report ships at full strength or the cell is theater.
• Scenario fiction without a decision attached → frame the decision first; the cell serves a choice.
• Point probabilities from nowhere ("Scenario B: 35%") → base-rate bands or honest ranking; never decorative precision.
• Skipping the reflexivity check → the report's own publication is an intervention; ask which futures the recommended action feeds.
• Scenarios without early indicators → futures you can't see arriving are futures you can't act on; every scenario gets watchable tripwires.
• Treating the cell's output as a verdict on people or projects → the precrime moral; forecasts inform, humans decide, nothing unhappened convicts.
• Re-running the cell forever instead of deciding → the log and the next review trigger close the loop; deciding is the deliverable.

Tailor to your environment

Record in references/your-environment.md: the standing decisions this cell serves (and their horizons), your reference classes and where their base rates live, who watches which tripwires, and where scenario decks and decision logs are filed. Sensitive specifics go in *.private.md (git-ignored).

Keep your filled-in copy outside the plugin. This file ships as a template and lives inside the installed plugin, where a /plugin marketplace update can overwrite it or refuse to run against a dirty tree. Copy it into your own project — .claude/skills-env/minority-report.private.md works well — fill it in there, and point this skill at that copy. Your specifics then survive updates and stay somewhere you own rather than in a cache you may not realise is disposable.

References

• references/scenario-cell-method.md — the full protocol: variable sorting, scenario construction with a worked domain-neutral example, the minority-report rule, the one-variable turn table, reflexivity patterns, probability discipline, tripwire design, and the decision-log template
• references/your-environment.md — your decisions, reference classes, watchers, filing (fill in)