scenario-branch
Skill0 starsChanged 4 months agoMITView on GitHub
Replace a single point-forecast with 3–5 internally consistent futures built from named driver variables, each with triggering conditions, an ordinal plausibility, and how it propagates to other domains. Use whenever forecasting, planning under uncertainty, stress-testing a strategy, assessing risk, or answering "what will happen with X." Triggers on "what's the outlook for," "where is this heading," "will X happen by year N," roadmap and contingency planning, red-teaming a bet, and any question about the future where a single confident number would be a lie. Especially apt when the honest answer is "it forks."
In this repository11 other files it gives its agents.
- analogy-checkSkill
- change-my-mindSkill
- disaggregateSkill
- disparate-impact-auditSkill
- grade-the-evidenceSkill
- hype-checkSkill
- prose-surgeonSkill
- protect-the-skillSkill
- steelmanSkill
- validate-the-claimSkill
- value-framesSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 34
Ask: "what are the few things that, if I knew their values, would tell me which future we're in?" Good drivers are:
- Distinctive: line 55
If you're tempted to give a confident single number past ~3 years, that's the signal to branch instead.
- Distinctive: line 119
Branches must be mutually distinguishable by their triggers: If you can't say which one you're in from observables, they're not real …
≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Scenario Branch
Presenting one future for a fast-moving, high-agency situation is itself a misrepresentation — it hides that the outcome depends on choices and unknowns that haven't resolved yet. But the usual fix, "optimistic / pessimistic / base case," is barely better: those labels describe your mood, not the mechanics. This skill builds real branches from the variables that actually drive the fork, so each future is a testable if-then, not a vibe.
What separates this from bad scenario work
- Branches come from driver variables, not moods. Identify the 2–4 variables whose values actually determine the outcome (pace of a capability, speed of a policy response, whether a cost curve keeps falling). The branches are combinations of those values — not "good case / bad case."
- Each branch names its trigger. State the observable condition under which this branch is the one you're in. A scenario you can't tell you're entering is useless.
- Plausibility is ordinal, not fake-precise. Rate each branch
central/plausible/tail. Do not attach invented probabilities ("23% likely") to things you cannot possibly have measured — false precision manufactures authority you haven't earned. - Branches propagate. Name which other domains each branch pushes on, through what mechanism, with what lag. A scenario that changes nothing downstream isn't a scenario, it's a footnote.
- Never average the branches. The expected-value mush across incompatible futures is a number no future will resemble. The set of branches is the answer; collapsing it destroys the information.
Driver variables: how to find them
Ask: "what are the few things that, if I knew their values, would tell me which future we're in?" Good drivers are:
- Uncertain (you genuinely don't know the value) — a variable everyone agrees on isn't a driver, it's a constant.
- High-leverage (its value swings the outcome).
- Trackable (you can watch it move — that's what makes triggers work).
Two to four drivers is the sweet spot. One driver gives you a line, not a fork. Five-plus gives you an unreadable combinatorial mess — fold the weak ones into assumptions.
Timescale discipline
The further out, the wider the fork, and the less a "forecast" is even the right object. Scale your register to the horizon:
- 0–1 yr: near-measurement; empirical evidence anchors directly.
- 1–3 yr: branches tighten around known drivers; still forecastable.
- 3–7 yr: scenario-dominant. State the capability/policy assumption each branch rests on, out loud.
- 7–15 yr: conditional statements only ("if driver X holds, then…"), not point forecasts.
- 15+ yr: possibility space, not prediction. Enumerate what could happen; refuse to say what will.
If you're tempted to give a confident single number past ~3 years, that's the signal to branch instead.
Procedure
- State the question and its horizon.
- Identify the 2–4 driver variables (uncertain × high-leverage × trackable).
- Build 3–5 branches from meaningful combinations of driver values. Give each a short evocative name and a one-paragraph internally-consistent story.
- For each branch: trigger (the observable that says you're in it), plausibility (central/plausible/tail), propagation (which domains it pushes on, mechanism, lag).
- State what to watch — the leading indicators that will tell you which branch is materializing (hands off to change-my-mind for the update triggers).
- Give a decision-useful takeaway: what's robust across all branches (do this now), and what to keep optionality on (decide later, when a driver resolves).
Output template
QUESTION: <what> | HORIZON: <window>
DRIVERS: <v1>, <v2>, <v3> (uncertain, high-leverage, trackable)
BRANCH A — "<name>" [central | plausible | tail]
Driver values: <v1=…, v2=…, …>
Story: <one internally-consistent paragraph>
Trigger (you're here if): <observable>
Propagates to: <domain> via <mechanism>, lag <immediate|months|years>
BRANCH B — "…" [ … ] … (repeat for 3–5)
WATCH: <leading indicators that discriminate between branches>
ROBUST NOW: <what to do regardless of branch>
OPTIONALITY: <what to keep open until a driver resolves>Example (compressed)
Question: Will AI broadly raise wages this decade? Horizon: 3–10 yr.
DRIVERS: (1) alignment/oversight matures before capability gaps? (2) does
governance distribute or concentrate the gains? (3) does compute access widen?
BRANCH A — "Electrification redux" [plausible]
Oversight matures + gains distributed. Lag, then broad wage participation —
like electricity's delayed but eventually shared payoff.
Trigger: rising labor share + falling entry-level displacement by ~year 3.
Propagates to: Education (pipeline stabilizes), Governance (less welfare stress), lag: years.
BRANCH B — "Engels' Pause" [central]
Capability works, gains captured by capital. Productivity up, median wage flat
for years — the 1800s pattern, until bargaining power shifts.
Trigger: productivity-wage divergence persists past year 3; labor share keeps falling.
Propagates to: Governance (redistribution pressure), Epistemics (AI-blame narratives), lag: immediate → years.
BRANCH C — "Great Divergence hardens" [plausible]
Compute stays concentrated in a few firms/regions. Rentier structure; gains
accrue to compute owners; Global South falls further behind.
Trigger: compute access gap widens; frontier stays in <5 firms.
Propagates to: Geopolitics (compute as leverage), Demography (migration), lag: years.
WATCH: labor share, entry-level hiring index, compute-access distribution, policy response speed.
ROBUST NOW: invest in retraining infrastructure + measurement — pays off in every branch.
OPTIONALITY: hold big redistribution-policy bets until the labor-share signal resolves (~year 3).Gotchas
- Branches must be mutually distinguishable by their triggers. If you can't say which one you're in from observables, they're not real branches.
- Don't smuggle your preferred future in as "central." Plausibility ratings are about mechanism and base rates, not what you're rooting for.
- 3–5 branches. Two is a false binary; six is unreadable. Fold weak drivers into stated assumptions.
- Pairs with change-my-mind (the triggers become update conditions) and grade-the-evidence (each driver's current value has an evidence tier).
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

