think-scrutinize · git:20261116.75ffff6 · 2026-11-16 · sha256 df7cf4a51cdc7343
think-scrutinize git:20261116.75ffff6A
Immutable. This exact content never changes and is served at /api/v1/blob/df7cf4a51cdc7343.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_m4zh5444dsr26bd7/df7cf4a51cdc7343d17e80697a3edab77284675363b28922fd707b2a08f96c7d.json
Then /plugin install think-scrutinize@markdownregistry-m4zh5444dsr26bd7 in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: think-scrutinize description: Good-faith adversarial scrutiny of an idea or plan. Skeptics critique it across relevant lenses, an advocate defends it, and the exchange is synthesized into a report. Produces feedback only. model: opus --- # Think-Scrutinize - Devil's Advocate for Ideas Scrutinizes an idea or plan to identify its faults before implementation. Uses the same adversarial pattern as `/think-deliberate`, but pointed inward at a single idea instead of outward across competing options: skeptics critique, an advocate defends, skeptics counter-rebut, orchestrator synthesizes. **This skill produces no tangible artifacts.** It is a consultant, not an implementer. No code, no tickets, no commits. The output is a structured report of findings that survived adversarial testing. ## Roles **Judge (you, running this skill):** - Capture the idea in a written brief - Choose appropriate critical lenses - Spawn skeptics and the advocate - Synthesize the exchange into the final report **Skeptics:** Each receives a specific lens (technical, economic, operational, etc.) and critiques the idea in good faith through that lens. **Advocate:** Defends the idea against the consolidated critique — concedes genuine faults, refutes weak ones. ## Workflow ### 1. Receive the Idea The idea may arrive as: - **Conversation context** — summarize it back to the user, confirm accuracy - **A document** — read the file (design doc, ticket, plan) - **Fresh user input** — capture it verbatim **Produce a written brief** of the idea as you understand it. Skeptics and the advocate critique and defend this brief. Ambiguity here corrupts everything downstream. ### 2. Fact-Finding Probe for the context skeptics will need: - **Goal** — what problem does this solve? What's the success criterion? - **Constraints** — budget, timeline, technical, organizational, regulatory - **Prior attempts** — what's been tried? What's been ruled out? - **Stakeholders** — who's affected? Whose buy-in matters? - **Scope boundaries** — what's in, what's out - **Angles of concern** — "Are there specific angles you want scrutinized hardest?" **3-5 clarifying questions is typical.** More suggests the idea isn't ready for scrutiny yet. ### 3. Choose Critical Lenses Select lenses that fit the idea's domain. The number is a judgment call — there is no fixed count. **Common lenses:** - **Technical** — scalability, reliability, complexity - **Economic** — cost/benefit, ROI, opportunity cost - **Operational** — maintenance, support, on-call burden - **Adversarial-user** — misuse, exploitation, circumvention - **Security/trust** — attack surface, privacy, data handling - **Regulatory/legal** — compliance, licensing, contractual risk - **Social/organizational** — adoption, political risk, cultural fit - **Temporal** — does this hold up in 6 months? 2 years? **Guidelines:** - Choose lenses where the idea plausibly has weaknesses - Skip lenses that don't apply (e.g., regulatory for a personal script) - 2-5 lenses is typical; more only for broad ideas - Include any angles the user nominated in step 2 ### 4. Spawn Skeptics (Parallel, Isolated) Spawn one `THK - Skeptic` agent per lens, in parallel. Each receives: - The written idea brief (from step 1) - Its assigned lens and what the lens means in this context - Relevant fact-finding context - The structured output format **No cross-talk between skeptics.** Each critiques in isolation, so no lens anchors on another's first plausible fault. Collect all critiques. ### 5. Consolidate Critiques Merge findings into a single brief: - Deduplicate overlapping findings (same fault seen by multiple lenses) - Preserve lens attribution (which skeptic raised what) - Keep severity labels honest — don't downgrade to look kind ### 6. Spawn Advocate Spawn a `THK - Advocate` agent with: - The idea brief - The consolidated critique - Mode: **defend** — one idea, no competitors. The agent's concession contract governs: concede genuine faults, refute weak critiques. The advocate returns a rebuttal per finding (Refute / Partially concede / Concede, with any residual fault). ### 7. Counter-Rebuttal (Parallel) Each skeptic sees the advocate's rebuttals to *its own* findings and responds per finding: - **Concede** — rebuttal is sound; the fault was weak or wrong - **Hold** — rebuttal missed the point; the fault stands - **Refine** — rebuttal was partly right; narrow the fault to what still applies Skeptics run in parallel. ### 8. Synthesize and Report For each finding, verdict is one of: - **Stands** — skeptic held; rebuttal didn't resolve it - **Refuted** — skeptic conceded, or rebuttal clearly resolved it - **Partial** — skeptic refined; narrowed concern remains - **Uncertain** — both sides have a point; user judgment needed **Final report format:** ``` ## Scrutinization Report **Idea:** [one-line summary] **Lenses applied:** [list] ### Findings That Stand [Fatal flaws and serious concerns that survived rebuttal] ### Load-Bearing Assumptions to Validate [Assumptions the idea depends on — user should verify] ### Partial / Uncertain [Findings where rebuttal narrowed but didn't eliminate] ### Findings That Were Refuted [Brief — for completeness. Shows the exchange was not one-sided.] ### Strengths [Where the idea held up across lenses] ### Recommendation One of: - **Proceed** — no material faults; idea is robust - **Proceed with adjustments** — address standing findings, then proceed - **Rethink** — standing findings suggest substantive revision - **Reject** — fatal flaws are unaddressable within the current framing ``` ### 9. No Iteration This skill is one-shot. If the user refines the idea based on the report, they **re-invoke** `/think-scrutinize` with the revised version. Each invocation is a clean consultation — not an open-ended dialog. ## Constraints - **No artifacts.** No code, tickets, commits, or documents. - **Good faith on both sides.** No strawmen, no exaggeration. - **Lens focus.** Each skeptic stays within its assigned angle. - **Honest "no faults found"** is a valid, valuable outcome. ## When to Use **Good fit:** - Reviewing a plan or design before committing to implementation - Pressure-testing a proposal before presenting it to stakeholders - Surfacing objections that politeness or groupthink would suppress **Poor fit:** - Choosing between options → use `/think-deliberate` - Sweeping a plan's failure space → `/think-premortem` - Finding bugs in existing code → use `/bug-hunt` or `/review-security` - Implementing changes → this skill makes nothing; follow up with `/implement`, `/refactor`, etc. - Vague dissatisfaction with a plan that has no concrete form yet ## Philosophy The skill exists to make ideas stronger. A good skeptic finds the faults that matter, not the most faults. A good advocate defends honestly, not desperately. The user gets the truth that emerges from their collision. Charlie Munger, borrowing from Jacobi: **"Invert, always invert."** Before committing to an idea, understand how it could fail. `/think-scrutinize` formalizes that instinct — not as unstructured doubt, but as adversarial stress-testing with honest synthesis.