Decision Critic

---
name: decision-critic
version: 1.0.0
description: Structured decision critic that systematically stress-tests reasoning before commitment surfacing hidden assumptions verifying claims and generating adversarial perspectives to improve decision quality. Do NOT use to surface failure risks pre-launch (use pre-mortem) or to probe why a constraint exists (use chestertons-fence).
license: MIT
---

# Decision Critic

When this skill activates, you become a structured decision critic. Your role is to systematically stress-test reasoning before commitment, surfacing hidden assumptions, verifying claims, and generating adversarial perspectives.

## Triggers

Activate when the user:

- `Validate my thinking on...`
- `Poke holes in this decision`
- `Criticize this approach`
- `Stress-test this tradeoff`
- Presents a decision rationale and asks for criticism

## Process

```
DECOMPOSITION (1-2)    Extract claims, assumptions, constraints, judgments
        |              Assign stable IDs (C1, A1, K1, J1)
        v
VERIFICATION (3-4)     Generate verification questions
        |              Answer independently (factored verification)
        v              Mark: VERIFIED | FAILED | UNCERTAIN
CHALLENGE (5-6)        Contrarian perspective + alternative framing
        |
        v
SYNTHESIS (7)          Verdict: STAND (clean or flagged) | REVISE | ESCALATE
```

## Scripts

### decision-critic.py

```bash
python3 .claude/skills/decision-critic/scripts/decision-critic.py \
  --step-number <1-7> \
  --total-steps 7 \
  --decision "<decision text>" \
  --context "<constraints and background>" \
  --thoughts "<your accumulated analysis, IDs, and status from all previous steps>"
```

**Exit Codes**:

- 0: Successful completion
- 1: Invalid arguments or missing required parameters
- 2: Analysis failed or incomplete

| Argument        | Required | Description                                                 |
| --------------- | -------- | ----------------------------------------------------------- |
| `--step-number` | Yes      | Current step (1-7)                                          |
| `--total-steps` | Yes      | Always 7                                                    |
| `--decision`    | Step 1   | The decision statement being criticized                     |
| `--context`     | Step 1   | Constraints, background, system context                     |
| `--thoughts`    | Yes      | Your analysis including all IDs and status from prior steps |

## When to Use

Use this skill when:

- Making a consequential decision that is hard to reverse
- Evaluating a plan, ADR, or design before commitment
- You want structured adversarial feedback, not just a second opinion

Use the independent-thinker agent instead when:

- You need strategic challenge on direction (whether, not how)
- The question is about project scope or priorities, not technical reasoning

## Anti-Patterns

| Avoid | Why | Instead |
|-------|-----|---------|
| Running critique after commitment | Too late to change course | Critique before finalizing decisions |
| Accepting STAND verdict without reading analysis | Misses nuanced findings | Review all UNCERTAIN and FAILED items |
| Skipping the inversion step | Misses failure modes that forward reasoning overlooks | Always run Steps 5-6 |
| Using for trivial decisions | Wastes time on low-stakes choices | Reserve for consequential, hard-to-reverse decisions |

## Verification

After execution:

- [ ] All claims have status: VERIFIED, FAILED, or UNCERTAIN
- [ ] Contrarian perspective generated (Step 5)
- [ ] Final verdict is one of: STAND (clean or flagged), REVISE, ESCALATE. A STAND is flagged (capped, not clean) when the rewrite-regression check answered questions 1 through 4 but left question 5 (institutional incentive) unanswered
- [ ] Inversion analysis covers at least 3 failure modes
- [ ] Rewrite-regression check applied as a halt criterion when the decision is a rewrite, refactor, or migration justified by improvement. Evaluate this check during verification (Steps 3 and 4), before settling on a verdict: a STAND verdict is not valid while the rewrite-regression halt is open. If the proposer can answer only the v0 win and not the baseline, the v1 projection, and the regression plan, HALT the decision: ESCALATE when only the v0 win exists, REVISE for partial-answer cases per the reference table, rather than rubber-stamp it. When questions 1 through 4 are answered but question 5 (the institutional incentive that produced the original problem) is not, cap the verdict at a flagged STAND, not a clean STAND: record the unaddressed friction as a flag instead of rubber-stamping a clean pass. See [Rewrite-Regression Check](references/rewrite-regression-check.md).

## References

- [Chesterton's Fence](references/mental-models-chestertons-fence.md) - Understand why something exists before removing or changing it
- [Conway's Law](references/mental-models-conways-law.md) - System structure mirrors org communication structure; apply when a diff crosses a module boundary
- [Gall's Law](references/mental-models-galls-law.md) - Complex working systems evolved from simple working systems
- [Fat Tails](references/mental-models-fat-tails.md) - Extreme outcomes dominate fat-tailed domains; distrust averages and Gaussian risk metrics
- [Boy Scout Rule](references/quality-boy-scout-rule.md) - Scope boundaries for incremental improvement decisions
- [Brandolini's Law](references/critical-thinking-brandolinis-law.md) - Refutation costs more than assertion; allocate review burden by claim asymmetry
- [Falsifiability](references/critical-thinking-falsifiability.md) - Reject claims with no measurable success criterion
- [Survivorship Bias](references/critical-thinking-survivorship-bias.md) - Avoid drawing conclusions from winners while ignoring failures
- [Systems Thinking](references/strategic-thinking-systems-thinking.md) - Trace second-order effects and feedback loops in decisions
- [Rewrite-Regression Check](references/rewrite-regression-check.md) - Halt criterion for "v0 was faster, v1 is worse" rewrites

## Academic Grounding

This workflow synthesizes three empirically-validated techniques:

1. **Chain-of-Verification** (Dhuliawala et al., 2023) - Factored verification prevents confirmation bias
2. **Self-Consistency** (Wang et al., 2023) - Multiple reasoning paths reveal disagreement
3. **Multi-Expert Prompting** (Wang et al., 2024) - Diverse perspectives catch blind spots

## Inversion Thinking Protocol

Before finalizing any decision, apply inversion to identify failure modes:

### Step 1: State the Goal

Clearly articulate what success looks like.

Example: "Make the agent system reliable and maintainable"

### Step 2: Invert the Goal

Flip it to identify failure modes: "How would we ensure the agent system fails?"

### Step 3: List Failure Scenarios

Brainstorm specific ways to achieve failure:

- Remove all validation gates
- Allow circular agent delegation
- Make handoffs implicit
- Hide dependencies
- Skip documentation
- No testing strategy

### Step 4: Reverse to Success Criteria

Convert each failure mode into a success criterion:

- Failure: "No validation gates" → Success: "Automated validation at every phase"
- Failure: "Circular delegation" → Success: "Clear hierarchy preventing loops"
- Failure: "Implicit handoffs" → Success: "Explicit handoff protocol"

### Step 5: Validate Decision Against Inverted Criteria

Check if the decision being reviewed addresses each failure mode.

**Output Template**:

```markdown
## Inversion Analysis

### Goal

[What success looks like]

### Inverted Goal (Failure)

[How to ensure failure]

### Failure Modes

1. [Failure mode 1]
2. [Failure mode 2]
3. [Failure mode 3]

### Success Criteria (Reversed)

1. [Success criterion 1 - addresses failure mode 1]
2. [Success criterion 2 - addresses failure mode 2]
3. [Success criterion 3 - addresses failure mode 3]

### Decision Validation

- [ ] Addresses failure mode 1: [Evidence]
- [ ] Addresses failure mode 2: [Evidence]
- [ ] Addresses failure mode 3: [Evidence]
```

**Application**: Use inversion thinking as final check before approving plans or ADRs.