Council of High Intelligence is a coding-agent skill that runs a structured deliberation over a hard decision. It assigns the question to a panel of deliberately different analytical personas, keeps their opening positions independent, forces cross-examination, and returns a verdict that leads with what remains unresolved, records dissent, and pairs the recommendation with kill criteria and one concrete next step.
It installs as a /council command in Claude Code, Codex, Gemini CLI, and OpenCode. The protocol targets decisions with material downside, competing values, incomplete evidence, or irreversibility, and its own guidance points to a direct answer, an experiment, or a smaller mode when a full council would not change the outcome.
| Mode | Shape | Use it when |
|---|---|---|
| Full | Independent analysis, cross-examination, final stance, synthesis | Stakes are high and competing frames need contact |
| Quick | Restate, rapid analysis, final positions | The decision needs breadth but not a full adversarial round |
| Duo | Opening positions, direct response, final statements | One polarity defines the decision |
The 18 members are analytical instruments rather than impersonations: each has a grounding protocol, a method, known blind spots, and a structured response contract, and each is paired with a counterweight (Aristotle with Lao Tzu, Socrates with Feynman, Torvalds with Meadows, Karpathy with Sutskever, and so on). Named triads such as architecture, risk, shipping, or ai-safety pick three lenses without convening the whole panel.
When several model providers are available, a detection script finds them (native host subagents, Codex, Gemini, Ollama, NVIDIA NIM, Cursor) and the coordinator spreads seats across them, separating polarity pairs so a single model family does not argue both sides. A --dry-route flag previews the seat map without running a council.
Features
- Three modes: full, quick, and duo, selected with --full, --quick, and --duo
- 18 persona contracts: stored under agents/, each with method, blind spots, and a response contract
- Panel selection: --members, named triads across 17 domains, and classic, exploration-orthogonal, and execution-lean profiles
- Round budget and enforcement: restate, blind analysis, cross-examination, final stance, synthesis, with checks for premature agreement, repeated claims, missing dissent, and unsupported confidence
- Evidence labels: FACT, INFERENCE, ASSUMPTION, and UNKNOWN mark every claim in a verdict
- Split verdicts: a divided weighted tally is returned as a split rather than flattened into consensus
- Outcome ledger: record prediction, owner, review date, and disconfirming evidence, then mark the result confirmed, revised, reversed, or inconclusive
- Multi-provider routing: seats distributed across detected providers, an explicit seat map through --models, and --no-auto-route to stay on the host
- Multi-host installer: one install.sh with per-host flags, custom target directories, and a dry-run mode
- Validation checklist: a script that checks persona structure, host-protocol parity, routing configuration, and verdict fields
