← ClaudeAtlas

conjecture-criticismlisted

Adversarial analysis via conjecture and criticism. Spawns parallel agents to independently generate competing approaches, then each agent critiques all others. Consensus emerges from cross-criticism. Use when making recommendations with real alternatives, evaluating options, or when the user says "analyze", "evaluate", "criticize", "compare options", "what am I missing", or "poke holes".
mrap/hex · ★ 2 · AI & Automation · score 78
Install: claude install-skill mrap/hex
# Conjecture & Criticism ## Philosophy Knowledge is created through conjecture and criticism, not authority or induction (David Deutsch, The Beginning of Infinity). Every recommendation is a conjecture. Every conjecture deserves genuine criticism from independent perspectives. The strongest idea is the one that survives. ## When This Activates ### Auto-trigger (baked-in via standing order #16) - The agent is about to recommend one option over alternatives - The domain has real consequences (architecture, strategy, process, tool selection) - Multiple viable approaches exist ### Manual trigger - User invokes `/analyze`, `/evaluate`, or `/criticize` - User asks "what am I missing", "poke holes", "compare options" ### Skip (do not trigger) - Tactical execution with a single clear path - Questions with factual answers - Tasks where the user has already decided and wants execution ## Depth Tiers | Tier | When | Agents | Rounds | Output | |------|------|--------|--------|--------| | Quick | Low stakes, tactical | 0 (inline self-check) | 1 | Verdict with brief "considered X but Y" | | Moderate | Multiple viable alternatives, real trade-offs | 3 | 2 (generate + all-to-all critique) | Verdict-first + key criticisms | | Deep | Architecture, strategy, process, or significant disagreement | 4+ | 3+ (generate + all-to-all + targeted debate) | Verdict + scorecard + killed conjectures | **Escalation rule:** If moderate-tier agents disagree significantly after cross-critique (no conv