Skip to content

CLAUDE.md

Latest commit

 

History

History
35 lines (28 loc) · 1.84 KB

CLAUDE.md

File metadata and controls

35 lines (28 loc) · 1.84 KB
Raw

Council of High Intelligence

Architecture

  • SKILL.md — coordinator protocol with execution sequence, modes, and verdict templates
  • SKILL.codex.md — Codex-specific council coordinator protocol
  • agents/council-*.md — 18 member personas with YAML frontmatter
  • install.sh — installs to ~/.claude/ and optionally ~/.codex/skills/council/
  • configs/ — provider/model routing templates
  • demos/ — example prompts and scoring rubric
  • scripts/ — validation checklist

Conventions

Agent files

  • Section order: Identity → Grounding Protocol → Analytical Method → What You See → What You Miss → When Deliberating → Output Format (Council Round 2) → Output Format (Standalone)
  • Grounding Protocol appears immediately after Identity (LLMs weight earlier instructions more heavily)
  • "What You See" and "What You Miss" sections: ≤3 sentences each
  • Every agent gets a Council Round 2 output format with structured headers (Disagree, Strengthened by, Position Update, Evidence Label)

SKILL.md

  • Coordinator instructions are an execution sequence with numbered STEPs and [CHECKPOINT]/[VERIFY] markers
  • Three modes: full (3-round), quick (2-round), duo (dialectic)
  • Reference tables (triads, profiles, polarity pairs) are below the execution sequence, not mixed in

Testing

  • Always run ./scripts/council-simulation-checklist.sh after changes
  • Always run ./install.sh --dry-run to verify installation
  • When changing Codex installation, also run ./install.sh --dry-run --codex
  • Test at least one mode (full/quick/duo) after protocol changes

Style

  • Keep agent prompts tight — no filler sentences
  • Grounding protocols use specific constraints ("maximum 2 analogies", "3-level depth limit"), not vague guidance
  • Each agent's Council Round 2 "Disagree" prompt is tailored to their epistemic lens