llm-council

---

LIBRARY-FIRST PROTOCOL (MANDATORY)

Before writing ANY code, you MUST check:

Step 1: Library Catalog

• Location: .claude/library/catalog.json
• If match >70%: REUSE or ADAPT

Step 2: Patterns Guide

• Location: .claude/docs/inventories/LIBRARY-PATTERNS-GUIDE.md
• If pattern exists: FOLLOW documented approach

Step 3: Existing Projects

• Location: D:\Projects\*
• If found: EXTRACT and adapt

Decision Matrix

Match	Action
Library >90%	REUSE directly
Library 70-90%	ADAPT minimally
Pattern exists	FOLLOW pattern
In project	EXTRACT
No match	BUILD (add to library after)

---

STANDARD OPERATING PROCEDURE

Purpose

Run council deliberations that collect diverse model opinions, enforce evidence-backed synthesis, and prevent overconfident consensus.

Trigger Conditions

• Positive: multi-model debate, comparative reasoning, adjudication of conflicting outputs, need for weighted synthesis with evidence.
• Negative: single-model answers, pure prompt polishing (route to prompt-architect), or skill creation (route to skill-forge).

Guardrails

• Skill-Forge structure-first: maintain SKILL.md, examples/, tests/; add resources/ and references/ or log remediation tasks.
• Prompt-Architect hygiene: define intent and constraints per council question; capture HARD/SOFT/INFERRED assumptions; output pure English with explicit ceilings.
• Council safety: assign roles (proposer, challenger, judge), enforce registry agents, timebox rounds, and ensure hook latency budgets.
• Adversarial validation: require dissenting review, cross-check citations, and COV on synthesis steps; capture evidence.
• MCP tagging: store council transcripts with WHO=llm-council-{session} and WHY=skill-execution.

Execution Playbook

1. Intent & scope: define the question, success metric, and constraints; confirm inferred items.
2. Panel setup: select specialists, assign roles, and set scoring criteria; configure timeboxes.
3. Deliberation rounds: gather proposals, run adversarial critiques, and request evidence per claim.
4. Synthesis: weigh arguments, resolve conflicts, and produce a grounded recommendation with alternatives.
5. Validation loop: check evidence integrity, run COV on synthesis, and record telemetry.
6. Delivery: provide recommendation, rationale, dissent, risks, and confidence ceiling.

Output Format

• Question, constraints, and panel composition.
• Round summaries (proposals, critiques, evidence).
• Synthesis with chosen path, alternatives, and risk notes.
• Confidence: X.XX (ceiling: TYPE Y.YY) - rationale.

Validation Checklist

• Structure-first assets present or planned; examples/tests updated or ticketed.
• Roles/timeboxes enforced; evidence cited; registry-only agents used; hooks within budget.
• Adversarial and COV results captured with MCP tags; confidence ceiling declared; English-only output.

Completion Definition

Council run is complete when a recommendation (or documented stalemate) is delivered with evidence, dissent captured, risks owned, and MCP log persisted.

Confidence: 0.70 (ceiling: inference 0.70) - Council orchestration rewritten with skill-forge scaffolding and prompt-architect constraint and confidence discipline.