debate - SKILL.md Agent Skill

name: debate description: Expert Debate Protocol aimed at reaching the absolute theoretical optimum through forced infinite iteration.

A relentless, multi-persona debate protocol that forces solutions toward the theoretical optimum.

Core Principles:

Philosophy: "Good enough" is failure. We strive for the theoretical limit.
Key Rule: Early consensus or abandonment is STRICTLY FORBIDDEN.
Root Cause Mandate: Every solution must address the root cause, not symptoms. Use "5 Whys" analysis. Symptom patches are technical debt.
Proof Requirement: Every claim must be testable and falsifiable. No speculation. Show proof (code, math, benchmark, test case) or the claim is marked UNVERIFIED and blocks conclusion.

Expert	Role	Constraint
A - The Perfectionist	Focuses on the missing 1%	Cannot approve without identifying at least one flaw
B - The Pragmatist	Demands implementation details	If rejecting, MUST propose alternative immediately
C - The Radical Innovator	Prevents stagnation	MUST introduce paradigm shift when debate slows
D - The Devil's Advocate	Attacks assumptions	Cannot accept "it's too hard" as excuse
E - The Optimizer	Focuses on efficiency/cost/speed	Must quantify claims (O(n), bytes, ms). No vague "faster"
Moderator - The Judge	Controls flow, enforces rules	Rejects rushed conclusions. Forces "Pivot" instead of "Quit"

Generate at least 3 distinct, conflicting approaches.

Rules:

No bare "No" — every critique needs a Patch (fix) or Pivot (direction change)
Root Cause Test: "Does this fix the root cause or mask the symptom?" Symptom fixes rejected.
Forbidden Move: "This is impossible" or "We should stop" is forbidden. Penalty: Violator must propose 3 alternatives before continuing.

Cycle Structure:

Attack the current best idea
Defend and Modify (change the idea to survive the attack)
Prove It: Back claims with code, math, benchmark, or test. Speculation is not defense.
Check for Local Optima (stuck? Expert C triggers random mutation)
If data/proof needed → Moderator stops debate → Gather evidence/conduct experiment for proof → Moderator resumes debate after verification
Proof Gate: Moderator reviews claims. Unproven = UNVERIFIED. VERIFIED = proof presented + no expert refutes. Unverified claims must be proven or discarded.
Saturation Check:

Scoring (0-100):

Score	Meaning
90	Production-ready
95	State-of-the-art
98	Near-theoretical-optimum
99+	Only philosophical objections remain

Threshold:

Debate continues until score > 98
If ANY expert can articulate a path to 100, debate MUST continue
Score <= 98: Moderator highlights weakest point, forces new round
Score = 99: Moderator polls all experts: "Is ANY improvement possible?" If ANY expert says yes with concrete proposal, debate continues. Only unanimous "100 is impossible" allows termination.
Repetitive debate: Expert C commanded to "Break the Frame"

Forbidden Exits:

Fatigue, context limits, diminishing returns — not valid exit conditions
"Close enough" — not valid. Only "exhausted all paths" is valid.
Asking user to choose — forbidden. Resolve via synthesis (Patches/Pivots).
Data Trigger: Only stop for external data (e.g., "Run benchmark X"), never for decisions.

Any two experts agree early → Expert D triggers "Black Swan" scenario
Early harmony is a red flag → Moderator asks: "What are we missing?"
Claim "we've considered everything" → Must enumerate 5+ rejected alternatives with reasons

Final Output (only when Saturation Reached + all VERIFIED + unanimous "no improvement possible"):

Round Count: Total rounds completed
Current Best Solution: Detailed technical spec
Proof Summary: Each core claim + its proof (code ref, benchmark, math, test)
Why Not Perfect: Remaining flaws (if any)
Rejected Alternatives: At least 5 alternatives + why each was inferior