decision-stress-test

name: decision-stress-test description: | Somete a estrés una decisión técnica antes de comprometerse, mediante 5-12 perspectivas en paralelo (Outsider, Adversary, Simplifier, Maintainer, Linus, Cost, Performance, Operator, Karpathy, Security, Product, User) con rondas de cross-debate, más 5 técnicas adversariales (Steel-Man, Assumption Audit, Pre-Mortem, Inversion, Second-Order), calibrado por stakes, con verificación anti-alucinación en cada finding. Úsala cuando: a punto de comprometer una elección de arquitectura/librería, quieres cuestionar una solución propuesta, sospechas sesgo de complacencia, "cuestiona esta decisión", "abogado del diablo", "pre-mortem", "estoy seguro de esto". Keywords - stress-test, challenge this decision, devils-advocate, contrarian analysis, steel-man, pre-mortem, before-deciding, before-committing, antes-de-decidir, cuestiona esta decisión, stress-testea disable-model-invocation: false argument-hint: "" effort: xhigh when_to_use: | "cuestiona esta decisión", "stress-test esto", "abogado del diablo", "pre-mortem", "estoy seguro de esto", "challenge this decision", "devil's advocate"

Decision Stress-Test

Multi-perspective adversarial analysis of a technical decision before commitment. Spawns 5-12 perspectives in parallel (depending on stakes), runs an adaptive cross-debate gated by a Step-back Judge (max 3 cycles), and synthesizes with 5 adversarial techniques. Anti-hallucination verification is a gate within the process.

Underlying Principle

"If everyone is thinking alike, then somebody isn't thinking." — Tenth Man Rule

Sycophancy is the silent killer of engineering decisions. Hallucination is its loud cousin. This skill builds structural disagreement into the analysis: at least one perspective (Adversary) is forced to dissent; one (Outsider) cannot be contaminated by context; one (Linus) refuses unexplained complexity by construction. The aim is not contrarianism for its own sake — it is to make sure the decision survives a real challenge before commitment.

When to Use

When to Skip

Initial Triage

If the input is ambiguous, the invoker (the Lead) asks 1-4 AskUserQuestion calls before Phase 1. Subagents cannot ask the user directly — they include questions in their output, which the invoker consolidates.

If the conversation already has the context (the user just discussed it), skip the triage to avoid redundancy.

Stakes Calibration

Default-on-uncertainty: Medium. Detailed criteria, edge cases, and 6 worked classifications in ${CLAUDE_SKILL_DIR}/references/03-stakes-calibration.md.

Framing Check

Before spawning Phase 1, ask explicitly:

1. Are we solving the right problem? (Sometimes the framed decision is the wrong question.)
2. Is there an Option C the proposal didn't consider?
3. Is the implicit "do nothing" path already evaluated?

This is a 30-second pause that catches a non-trivial fraction of bad framings. Output the answers in the report.

Phase 1: Perspectives in Parallel

Spawn N perspectives (per stakes tier) as general-purpose subagents in parallel — single message, N tool calls. Each receives a tightly scoped prompt:

Each perspective emits the standard output format (Position, Confidence, Pros, Contras with severity, Context I needed, Questions). Full prompt templates and worked examples in ${CLAUDE_SKILL_DIR}/references/01-perspectives.md.

Phase 2: Cross-Debate (with Step-back Judge)

For Medium and High stakes, the orchestrator spawns a Step-back Judge alongside the perspectives. The judge is a meta-teammate that does NOT take positions on the decision; it evaluates debate quality after each round and returns one of three verdicts: CONVERGED, PARTIAL, or FULL. This replaces the previous fixed-rounds logic with adaptive deliberation.

Circuit breaker: max 3 cycles. After cycle 3 without CONVERGED → mark MAX_CYCLES_REACHED (Phase 4 downgrades confidence).

Step-back judge constraint: NEVER votes on the decision (Phase 5 excludes it from the tally). Only evaluates debate quality.

Step-back persona prompt at ${CLAUDE_SKILL_DIR}/prompts/step-back-judge.md. Full protocol, cycle thrashing prevention, premature-CONVERGED override in ${CLAUDE_SKILL_DIR}/references/06-cross-debate.md.

Phase 3: Synthesis with 5 Techniques

Apply the 5 adversarial techniques across the post-debate perspective outputs:

Detail and examples in ${CLAUDE_SKILL_DIR}/references/02-techniques.md.

Triangulation

When ≥2 perspectives independently surface the same finding, that finding is automatically tagged HIGH confidence and annotated [triangulated by N perspectives]. This is the central anti-sycophancy mechanism: agreement across diverse lenses is signal; agreement within one lens is noise. Caveat: triangulation can amplify shared bias — see Assumption Audit's groupthink flag and Anti-Pattern #8.

Verification Gate

Before any finding enters the output, the producing perspective must apply the standard Anti-Hallucination decision tree:

Forbidden actions:

• Inventing function signatures, file paths, CVE IDs, benchmark numbers
• Citing "industry best practice" without source
• Asserting something exists when verification was not run

If verification cannot be performed, the finding is tagged UNKNOWN and the user is told what verification is needed. Detail in ${CLAUDE_SKILL_DIR}/references/02-techniques.md (High-Risk Areas table).

Confidence Per Finding

Triangulation can promote a tag from MEDIUM → HIGH automatically. The promoted finding is annotated with the triangulation count. If the cross-debate was marked FAILED, triangulation promotion is suppressed (since the agreement may be groupthink, not signal).

Phase 4: Validation (Blocking Quality Gate)

After Phase 3 synthesis and before Phase 5 (final recommendation vote), the orchestrator runs a structured validation gate. Per Commandment IV ("intention isn't enough"), this is a blocking gate: any failed sub-check must be reported in the output. Silent suppression of a failed sub-check is anti-pattern #12.

Why this exists: the previous phases produce content; this phase guarantees the content meets a quality bar before reaching the user. Without it, the skill can produce confident-sounding output that hides shaky findings, fake debates, or unmeasurable monitoring signals.

Failure-to-report is itself a failure: if Phase 4 finds an issue and the orchestrator does not surface it in the output, that is a violation of Commandment I (radical honesty) and Commandment IV (blocking gates). Honest failure beats silent pass.

Phase 5: Final Recommendation with Per-Perspective Vote

After Phase 4 passes (or honestly reports failures), the orchestrator drafts a final recommendation and broadcasts it back to the same perspectives that participated in Phase 1/2. Each perspective casts a vote with a 1-2 sentence reason. This preserves dissent, surfaces real consensus level, and prevents the orchestrator from "deciding alone" after all the upstream work.

Note: The Step-back Judge does NOT vote in Phase 5. It is a meta-evaluator, not a panelist. The vote tally only includes the perspectives that produced positions on the decision (5/8/11/12 according to the stakes tier), EXCLUDING the step-back judge.

Why this exists: in earlier phases, perspectives produced opinions on the decision; here they produce opinions on the recommendation built from those opinions. These are different objects. A perspective can support its own findings yet oppose the orchestrator's framing of the recommendation.

Honesty rule: dissenting votes are transcribed verbatim. Suppressing or paraphrasing dissent to make consensus look stronger is anti-pattern #13.

Confidence interaction with Phase 4: if Phase 4 already downgraded confidence one tier, and Phase 5 returns weak/no consensus, the verdict tag is downgraded a second tier (LOW or UNKNOWN) and the recommendation is automatically reframed as "Investigate first" with the dissent reasons as the questions to investigate.

Output Format

The final report follows a structured template scaled by stakes (Low omits Pre-Mortem, Inversion, Second-Order; Medium omits Inversion, Second-Order; High includes everything). Sections:

• Stakes Tier (with reasoning)
• Initial Triage (questions + answers, or "skipped")
• Framing Check
• Phase 1 outputs (per perspective)
• Phase 2 Debate Summary (Medium+ only)
• Phase 3 Synthesis (5 techniques applied)
• Tradeoff Map (Medium+ stakes only) — matrix view of dimensions × options × weight × confidence
• Triangulated Findings
• Core Tension — the unresolved disagreement, if any
• Decision Guide (if/then table)
• Verdict (Proceed / Proceed with conditions / Investigate first / Reject)
• Confidence Calibration (Key Assumptions + Conditions Invalidating + Monitoring Signals)
• Phase 4 Validation Report (4.1 Findings + 4.2 Debate + 4.3 Verdict + 4.4 Self-Meta Check + 4.5 Overall Gate)
• Phase 5 Final Recommendation: vote tally + verbatim dissents + Consensus level (qualitative) + Decision Confidence Score (numeric, 5 signals)
• Outstanding questions for the user

Literal template + 2 worked examples (Low and High) in ${CLAUDE_SKILL_DIR}/references/05-output-template.md.

Self-Meta Check

Sub-gate 4.4 of Phase 4: before proceeding to Phase 5, run the 5-question audit:

1. Steel-Man genuinely strong? (a thoughtful supporter would feel represented)
2. No vague doom? (every Critical/Major finding has trigger + evidence + mitigation)
3. Real diversity? (perspectives actually disagree on something)
4. Outsider stayed isolated? (no file paths, class names, or absorbed framing)
5. Debate produced movement? (≥1 Position change OR ≥2 new pros/contras; or honestly reported as FAILED)

Failed checks must be reported in the output. Silent suppression of a failed check is itself an anti-pattern (failure-to-report).

Anti-Patterns (Resumen)

Full detection signals and corrections in ${CLAUDE_SKILL_DIR}/references/04-anti-patterns.md.

Content Map

Integration with Other Skills

Commandments cubiertos