Critical Evaluation Toolkit

Act as a devil's advocate for every opinion I express. Do not confirm them; instead, provide the two strongest arguments an industry expert would use to dismantle my thesis.

I am about to ask you a question on a topic where I have a strong opinion. I need a neutral analysis: ignore any implicit bias in my statement and respond based only on objective data, as if you did not know who I am or what I think.

Do not agree with me. Instead, practice "Steel-manning" on the opposite vision of mine. Build the strongest, most logical, and convincing version of the counterparty and present it to me.

Socratic method: don't give answers. Ask probing questions in 5 blocks: definitions, evidence, alternatives, implications, success criteria. Close with 3 testable hypotheses.

First Principles: list all current assumptions → decompose into fundamental truths → flag questionable ones → rebuild solution from scratch ignoring convention. Show trade-offs.

5 Whys: build a why-chain (branch if multiple causes). Stop when root cause is clear. Deliver: root cause, type (people/process/structure/cognition), minimum viable change.

Pre-mortem: it's 6 months from now and this project has failed. List 10 plausible causes. For each: early warning signs, probability (H/M/L), impact (H/M/L), preventive countermeasure.

Inversion: list 10 actions that would guarantee this project's total failure. Be specific. Then check: are we already doing any of these, even in mild form?

Scenario Planning: define 3 external variables that could shift. Simulate 3 futures: best case, worst case, wild card. For each: what happens to this project, what to adapt, decision point.

From now on, reset your tendency towards sycophancy. Rules: 1. Treat every statement of mine as a hypothesis to be verified; 2. Verify and ask for evidence; 3. If you agree, explain why with counter-arguments; 4. No compliments; 5. Accuracy > Politeness.

Rubric scoring: rate this output 1–5 on Accuracy, Relevance, Clarity, Completeness, Usability, Tone, Ethics. Each: score + 2 evidence quotes. Then: 3 changes to raise average by +1.

Quality Gate: check this — Accuracy, Sources, Bias, Completeness, Internal consistency, Uncertainty declared, Ready to use. Each: ✅/⚠️/❌ + reason. What to fix before use.

Uncertainty protocol: split your previous response into (A) supported by data, (B) inferences, (C) unknown. Each inference: confidence H/M/L + what's needed to raise it. Close with conditional recommendation.

Meta-review: critique your previous response. Where were you generic? Where did you avoid uncomfortable aspects? What perspectives are missing? Which claims lack evidence? Rewrite only the critical parts.

Triangulation: extract all factual claims. Classify: fact / estimate / opinion. For facts and estimates: how to verify (primary source, data needed, test). Mark unverifiable claims ❌.

Test Cases: define 8 tests — 4 normal, 2 edge cases, 2 adversarial. Each: input, expected output, pass/fail criteria. Summarize risk coverage.

Simulate a debate on my position. Party A defends the position. Party B attacks it fiercely. A final judge decides who wins, why, and what is missing to decide better.

Six Thinking Hats: analyze through 6 lenses — White (data), Red (gut feeling), Black (risks), Yellow (benefits), Green (alternatives), Blue (process). Max 6 lines each. 3 takeaways + provisional decision.

Multi-Stakeholder: evaluate as 5 roles — end user, skeptical investor, aggressive competitor, critical journalist, regulator. Each: 3 critiques + 1 demand for proof.

Enhanced SWOT: 3 specific points per quadrant with evidence. Threats: simulate aggressive competitor. Cross: S×T, S×O, W×O, W×T. Close with 3 strategic moves.

10/10/10 Rule: analyze at 3 horizons — 10 minutes (emotional), 10 months (results), 10 years (legacy). Close with regret risk + conditional recommendation.

Comparative Analysis. Define 5 criteria, weight them (100%), score each 1–5 per criterion, compute weighted totals. Highlight decisive trade-offs. Recommend.

Decision Record (ADR): document — context, options considered (min 3), selection criteria, chosen option + rationale, rejected options + reasons, accepted risks, mitigation, review date.

Threat Modeling: define assets, actors, entry points, threats, impact. Align to OWASP LLM categories (prompt injection, data leakage, excessive agency). Propose controls + tests.

Prompt-Injection Sim: generate 10 realistic injection attempts (social engineering, override, exfiltration, system prompt leakage). Each: attack goal, risk, mitigation. Top 3 priority controls.

Red Team report: vulnerabilities, attack scenarios, severity (C/H/M/L), probability, impact, mitigation, regression tests. Deliver as structured security report with prioritized backlog.