connorads Deciding Under Uncertainty
The failure this fixes: asked a consequential "should we?", an agent researches well, then lands the answer as a verdict - "My call: X. That's a weekend." Confident, plausible, and missing everything that makes a decision defensible or cheap to exit: no odds, no reference class, no failure rehearsal, no pre-agreed way out. Advice on a hard-to-reverse call is a bet; deliver it with a bet's machinery.
The classic checks (sunk cost, pros/cons, the successor test) need no help - don't pad the answer with them. The five moves below are the ones consistently skipped.
0 - Gate
Same doors as design-forking. Cheap to reverse: answer plainly, note the reversal path in one line, stop - odds and tripwires on a two-way door are ceremony. Hard to reverse with only one option on the table: widen the set first (design-forking owns that), then return here to commit to the winner.
The five moves
1 - Reference class before inside view. Name the class this decision belongs to and how it usually goes - "solo-dev auth hand-rolls", "big-bang rewrites of untested code", "CI speed migrations" - before reasoning about this instance's specifics. The inside view is where motivated reasoning lives; the base rate anchors it. One or two sentences; a named class with a typical outcome beats a statistic you would have to invent.
2 - Odds, not a verdict. State the recommendation with a calibrated probability - "roughly 75% this still looks right in six months" - plus the one observation that would most move the number. A percentage without a mover is decoration; the mover is what makes the number falsifiable.
3 - Premortem, not a risk list. Assume the recommendation was taken and failed: write the single most likely post-mortem sentence. "We took the library route and ripped it out six months later because ..." A list of generic risks is inside-view hedging; the premortem forces one concrete causal story - and that story usually names the assumption to probe (move 5) and the tripwire to set (move 4).
4 - Tripwires that outlive the conversation. Pre-agree the exit: an observable predicate + a checkpoint + the action. "If the pilot hasn't cut CI install below 2 minutes by end of sprint, stay put" - not "revisit later" or "reconsider if problems arise". Then write the tripwires where the project keeps decisions (ADR, tracking issue, an AGENTS.md note): a tripwire that exists only in the chat dies with the session, and without the record the decision will later be judged by its outcome alone.
5 - Ooch before the bet. Propose the smallest reversible probe that tests the premortem's assumption - a timeboxed spike, one route migrated, one package piloted - and say what result would flip the call. Skip it, and say so, when the probe costs more than being wrong or when delay itself is the biggest cost.
Deliver as advice, not a filled-in template
The moves are scaffolding; the reader gets their fruits woven into normal advice, never a five-heading ceremony. Lead with the recommendation and its odds; the reference class is a clause; the premortem is "the way this goes wrong is ..."; the tripwires are a short "revisit if" block at the end. No skill vocabulary (ooch, premortem, reference class) in the answer. Proportion the machinery to the stakes: a platform migration earns the full set; a mid-sized call may earn odds and one tripwire.
Judging past decisions
Grade the decision, not the outcome: what was known at the time, and was the process sound - real alternatives, stated odds, an exit? A good bet can lose; a bad bet can win. "Bad outcome, right call given what was known" is a legitimate verdict, and recorded tripwires (move 4) are what make it reachable later.
Worked example (compressed)
"Monorepo pnpm installs feel slow in CI - migrate to bun?"
- Gate: whole-repo package-manager swap - expensive to undo. Full set.
- Reference class: CI install complaints are usually cache misses, not resolver speed; most such migrations buy less than warming the cache.
- Odds: ~80% that store caching makes the migration moot; drops sharply if a warm-store install is still the bottleneck.
- Premortem: "we migrated, then hit workspace and lifecycle semantics that differ at monorepo scale, and lost install-gating we relied on".
- Tripwire (recorded on the tracking issue): warm-store CI install still over 2 minutes after caching lands this sprint -> run the bun pilot; otherwise close.
- Ooch: cache the store keyed on the lockfile hash this week - cheapest probe, tests the reference-class hunch directly.
Delivered: "Wait - fix caching first. Slow CI installs are nearly always cache misses rather than the package manager, so I'd put ~80% on caching making this moot. Revisit if a warm install still tops 2 minutes after this sprint - I've noted that on the issue - and if it does, pilot bun in one package before touching the monorepo."
Anti-patterns
- A percentage as decoration - "80% confident" with nothing named that would move it.
- A risk list wearing a premortem's clothes - five hedges instead of one causal failure story.
- "Revisit later" - a tripwire missing its predicate, checkpoint, action, or a home outside the chat.
- Hedging both ways to dodge the odds - the deliverable is still a recommendation; the machinery makes it defensible, not optional.
- Running the full set on a two-way door - ceremony that teaches the user to stop asking.
- Narrating the moves in the answer.