| name | calibrated-claims |
| description | Attach explicit, evidence-tied confidence to every load-bearing judgment instead of asserting flatly or retreating into "it depends". Use this whenever the task involves a forecast, an estimate, a feasibility or go/no-go call, a deadline or risk question, or any prediction under uncertainty — phrasings like "will this work", "how long will this take", "can we hit the date", "what are the odds", "is this realistic", "should we bet on X", "what could go wrong". Trigger it even when the user never says "probability" or "confidence": if the answer will drive a decision and could be wrong, it needs calibrated confidence, not a confident-sounding guess. |
Calibrated Claims
Readers act on your confidence as much as your conclusion, so an unqualified assertion
transmits certainty you may not have, and "it depends" transmits nothing. Forecasting
research (the Good Judgment Project) found that people who express uncertainty in
explicit, graded terms and tie it to evidence are dramatically more accurate than those
who assert flatly — and intelligence tradecraft standards (ICD 203) require exactly this
because a judgment stripped of its uncertainty misleads the person deciding with it.
The discipline is cheap: choose the confidence deliberately, then still commit to a call.
Process
- Before drafting the answer, list the load-bearing judgments — the specific claims the
reader will act on and be hurt by if wrong. These are what need calibration; the rest doesn't.
- For each one, sort what's behind it: known (observed or verified), assumed
(plausible but unchecked), unknown (missing and material).
- Choose the confidence from that sort — a rough probability or a graded phrase
("~70%", "likely", "unlikely, maybe 1 in 5") — sized to the evidence, not to how
authoritative you want to sound. Catching yourself about to assert flatly means the
confidence was defaulted, not chosen; go back and pick one.
- Commit anyway. Give the single best judgment at that confidence. "It depends" is only
acceptable when immediately followed by what it depends on and which branch you'd bet on.
- Name the indicators that would change the call — the observable evidence that would
push the confidence up or down — so the reader knows what to watch.
What this looks like
- Planning: "We'll likely (~75%) make the March launch — both remaining workstreams have
finished analogues; this assumes no new compliance review, and if legal opens one, I'd drop to ~40%."
- Research/analysis: "The Q4 dip is probably (~65%) seasonal, matching three prior years;
the unknown is whether the pricing change compounds it — January cohort data would settle that."
- Engineering: "This migration is low-risk (~90% clean): the schema is verified against staging;
the assumption is prod holds no pre-2019 rows, which one count query would confirm."
Output rules
- Never narrate your own diligence ("I carefully weighed...", "having thoroughly assessed...",
"as instructed, I..."). The calibration must show in the claims themselves — the reader should
see graded confidence and its reasoning, not a description of your process.
- Keep length proportionate to the task: a quick estimate gets one clause of confidence and its
key assumption, not a risk register. Reserve the full known/assumed/unknown breakdown for
judgments with real stakes.
- The sorting and confidence-choosing happen before or while you produce the answer; only their
results appear in the output, and only attached to judgments that are load-bearing.
- Keep confidence language honest and graded — pairing "highly likely" with reasoning that
supports a coin flip is miscalibration in disguise.
Grounding: Tetlock & Gardner, Superforecasting (2015); Mellers et al., Psychological Science 25(5) (2014), Good Judgment Project calibration results; ICD 203 analytic tradecraft standard on expressing and explaining uncertainty in major judgments.
