---
name: decision-retrospective
description: "Review a past decision against its original evidence and observed outcomes to improve decision rules."
---

# Decision Retrospective

Learn from a decision without rewarding luck, punishing bad luck, or rewriting history around the result.

## Boundary

Use this skill after a real decision has been made and there is an outcome, new evidence, or meaningful passage of time to examine.

- Use pre-decision analysis when the choice is still open.
- Do not audit a person’s character or assign motives from one decision.
- Do not treat an unfavorable outcome as proof of a poor process.
- Do not treat a favorable outcome as proof of a sound process.

Missing records narrow the verdict. They do not justify confident reconstruction.

## Required Inputs

Collect:

- decision, date, owner, authority, and affected stakeholders
- objective, decision criteria, constraints, and deadline
- options considered, including delay or no action where credible
- evidence and assumptions available at the time
- expected outcomes, risks, and confidence stated at the time
- decision process, dissent, communication, and planned review point
- actual outcome and material changes after the decision

Keep contemporaneous records distinct from later recollection.

## Workflow

### 1. Reconstruct the decision-time record

Before using the outcome to judge the process, document what was available at the time:

- facts and source quality
- unknowns
- forecasts and confidence
- constraints and time pressure
- credible alternatives
- stakeholder input and unresolved dissent

If the outcome is already known, actively check for hindsight edits such as “we always knew” or “it was obvious.” Prefer dated artifacts over memory.

### 2. Restate the decision model

Write the expected chain:

```text
chosen action -> expected mechanism -> expected outcome
```

Record the criteria used, assumptions required, rejected alternatives, anticipated downside, and conditions that would trigger review or reversal.

### 3. Audit the process

Assess:

- whether the problem and objective were framed correctly
- whether alternatives were materially different and honestly compared
- whether evidence quality matched the stakes
- whether relevant base rates or prior cases were considered
- whether authority and consultation were appropriate
- whether incentives, dissent, and inconvenient evidence were visible
- whether speed matched reversibility, cost of delay, and blast radius
- whether the decision and rationale were communicated clearly
- whether monitoring, rollback, and a review date existed when needed

Mark unsupported dimensions `Unknown` rather than failing them automatically.

### 4. Examine the outcome separately

Describe what happened, then separate:

- effects plausibly caused by the decision
- execution quality after the decision
- external changes
- randomness or favorable timing
- indicators that were missing or misread

Do not claim causation merely because the outcome followed the decision.

### 5. Compare expectation with reality

Identify calibration errors, omitted alternatives, assumptions that failed, signals that deserved more or less weight, and outcomes the original process could not reasonably have predicted.

Use counterfactuals cautiously. A plausible alternative outcome is not evidence that the alternative would have performed better.

### 6. Update future practice

Convert each material lesson into one of:

- changed decision rule
- evidence to collect earlier
- stakeholder or authority change
- reversible test
- monitoring or review trigger
- communication change
- no change, when the process was sound and the result was ordinary uncertainty

Name where the new rule should apply and where it should not.

## Verdict

Use one:

- **Sound process:** evidence supports the process despite the outcome.
- **Sound with gaps:** the process was defensible but had correctable weaknesses.
- **Weak process:** material avoidable defects undermined the decision quality.
- **Unverifiable:** the available record cannot support a responsible process verdict.

## Output

1. **Decision and evidence boundary**
2. **Decision-time model**
3. **Process findings** with evidence and impact
4. **Outcome analysis** separated from process quality
5. **Expectation versus reality**
6. **Updated decision rules and review triggers**
7. **Verdict and remaining unknowns**

Return up to five material findings and up to three changed decision rules. Zero findings or rule changes is valid when the evidence supports no change. Do not reproduce the full chronology unless it is needed to establish causation or the user requests it.

## Quality Gate

- The process is reconstructed from information available at the time.
- Hindsight is identified rather than used as evidence.
- Process quality and outcome quality remain separate.
- Findings cite records or are labeled Unknown.
- Counterfactuals are not presented as observed facts.
- Every lesson changes a future rule or explicitly justifies no change.