N-round opposing-stance debates for trade-off analysis. Assigns pro/con roles to agents, runs structured debate rounds with quality scoring, and produces a moderator synthesis with confidence-rated recommendation. Generalizable to architecture, technology, security, and design decisions.
Installer avec Codex ou Claude Copiez ce prompt, collez-le dans Codex, Claude ou un autre assistant, puis laissez-le vérifier la page du skill et l'installer pour vous.
Une commande directe contourne le prompt de vérification. Examinez la source avant de l'exécuter.
N-round opposing-stance debates for trade-off analysis. Assigns pro/con roles to agents, runs structured debate rounds with quality scoring, and produces a moderator synthesis with confidence-rated recommendation. Generalizable to architecture, technology, security, and design decisions.
["Assign clear opposing stances before debate begins — never let agents pick their own stance mid-round","Require each rebuttal to explicitly address the prior argument (no topic drift)","Score each round immediately before proceeding to prevent recency bias","Moderator must cite specific debate evidence in the final recommendation","Limit rounds to 3-5 to keep arguments focused; more rounds rarely change outcome"]
Runs a structured N-round debate between two opposing agents (pro and con) on any decision-requiring topic, then synthesizes a moderator recommendation. Built for architectural decisions, technology choices, security trade-offs, and design reviews — anywhere strong opinions exist and the right answer is non-obvious.
When to Use
Architecture decision records (ADRs) needing structured justification
Technology selection (e.g., Redis vs Postgres, REST vs GraphQL)
Security trade-off analysis (e.g., strict vs permissive auth policy)
Design review of competing approaches
Any HIGH/EPIC task where the planner identifies genuine disagreement among specialists
Do not use for:
Settled best practices (where there is consensus)
Simple implementation tasks
Decisions already made and documented in ADRs
Iron Law
NO RECOMMENDATION WITHOUT CITED DEBATE EVIDENCE
The moderator's synthesis MUST reference specific arguments made during rounds. A recommendation that ignores debate content is invalid.
Debate Protocol
Pre-Debate: Stance Assignment
Before any rounds begin, the invoking agent MUST define:
Topic — one clear decision question (not open-ended)
Pro stance — what the PRO agent advocates for
Con stance — what the CON agent argues against / alternative
Number of rounds — default 3, max 5
Success criteria — what a good recommendation looks like
Example:
Topic: Should we use event sourcing for the order service?
Pro stance: Yes — event sourcing provides audit trail and temporal queries
Con stance: No — event sourcing adds complexity; CRUD with snapshots is sufficient
Rounds: 3
Round Structure
Each round follows this fixed sequence:
Step 1: PRO Argument
PRO agent argues its stance with:
- 1 primary argument (specific, concrete)
- 1 supporting piece of evidence (metric, case study, or first-principles reasoning)
- 1 anticipated objection pre-addressed
Step 2: CON Rebuttal
CON agent responds with:
- Direct refutation of PRO's primary argument (must engage it — no deflection)
- 1 counter-argument from CON's stance
- 1 counter-evidence or challenge to PRO's evidence
Step 3: Round Score
Score each round immediately using this rubric (0–10 per dimension):
Dimension
Description
Specificity
Arguments use concrete examples, not vague claims
Evidence
Claims backed by data, precedent, or first principles
Rebuttal Quality
Directly addresses opposing argument, no deflection
After all rounds, the moderator agent produces a structured synthesis:
## Debate Summary**Topic:** [restate topic]
**Rounds completed:** N
**PRO total score:** X.X/10
**CON total score:** X.X/10
## Key Arguments**Strongest PRO argument:** [cite round + argument]
**Strongest CON argument:** [cite round + argument]
## Decision Factors
List 3-5 factors that should determine which stance wins in context:
1. Factor: [e.g., team expertise] → Favors: [PRO/CON/neutral]
2. ...
## Recommendation**Decision:** [clear recommendation — not "it depends"]
**Confidence:** [High / Medium / Low]
**Rationale:** [2-3 sentences citing specific debate evidence from rounds]
## Conditions / Caveats
[Any conditions under which the recommendation should be revisited]
Workflow
Step 0: Load Context and Memory
# Check for prior debate results on same topic
grep -r "adversarial-debate" C:/dev/projects/agent-studio/.claude/context/memory/decisions.md 2>/dev/null | tail -5
Check .claude/context/memory/decisions.md for prior ADRs on same topic. If a prior debate exists, acknowledge it and note whether conditions have changed.
Input validated against schemas/input.schema.json before execution.
Output contract defined in schemas/output.schema.json.
Pre-execution hook: hooks/pre-execute.cjs — validates topic, stances, rounds range (1-5).
Post-execution hook: hooks/post-execute.cjs — emits observability event via send-event.cjs.
Anti-Patterns
Never let agents pick their own stances — stance drift produces muddled debates
Never skip round scoring — unscored debates allow moderator bias in synthesis
Never produce "it depends" recommendations without explicit conditions listed
Never use more than 5 rounds — beyond round 3, new arguments rarely emerge
Never invoke debate for trivial decisions — overhead not justified for settled questions
Related Skills
llm-council — parallel multi-LLM synthesis (complementary: debate is sequential, council is parallel)
advanced-elicitation — meta-cognitive reasoning for single-agent analysis
plan-generator — use debate output as an ADR input to plan-generator
sequential-thinking — structured problem decomposition for the topic before debate
Memory Protocol
Before starting: Read .claude/context/memory/decisions.md for prior ADRs on related topics.
After completing: Append structured summary to .claude/context/memory/decisions.md with date, topic, recommendation, confidence, and key evidence citations.
Do not write directly topatterns.json, gotchas.json, or open-findings.json — use MemoryRecord tool.
