# Decision Critic > 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). - Skill: `rjmurillo/decision-critic` (Agent Skill, multi-file: 13 files) - Install (CLI): `npx skillmds@latest add rjmurillo/decision-critic` - Raw SKILL.md: https://api.skillmd.com/api/skills/rjmurillo/decision-critic/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - License: MIT - Author: rjmurillo (https://skillmd.com/u/rjmurillo) - Updated: 2026-09-10 - Page: https://skillmd.com/skills/rjmurillo/decision-critic --- # 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 "" \ --context "" \ --thoughts "" ``` **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.