Claude Council

Convene a council of independent Claude subagents to debate the user's question, then synthesize their perspectives into one plan. Every member runs in a fresh, isolated context (they cannot see this conversation or each other's first-pass answers), so you get genuinely independent takes — all billed as normal subagent usage on the user's subscription. No API keys are involved.

The three stages

1. Opinions — council members answer the question in parallel, each through a distinct persona lens.
2. Peer review (full mode) — each member receives all answers anonymized as "Response A/B/C/D", critiques them, and ranks them best→worst.
3. Chairman synthesis — you (the main session) de-anonymize, weigh the critiques and rankings, and present the final plan.

Default council

Why all Opus 4.8: as of mid-2026, Claude Code subagents cannot run Fable 5 — model: "fable" silently falls back to Opus 4.8, so Opus 4.8 is the smartest spawnable model. Member diversity comes from personas + fully isolated contexts; the chairman (main session) runs on whatever model the user's session uses. If Fable subagents become available, set Architect/Skeptic to "fable" — the script already accepts it.

How to run

Invoke the Workflow tool with the bundled script (runs in background; you'll be notified when done). The script council-workflow.js lives next to this SKILL.md — resolve its absolute path from wherever this skill is installed (user-level: ~/.claude/skills/claude-council/council-workflow.js; project-level: <project>/.claude/skills/claude-council/council-workflow.js).

Workflow({
  scriptPath: "<absolute path to council-workflow.js>",
  args: {
    question: "<the user's question, verbatim>",
    mode: "full",            // "full" (default) = with peer review; "quick" = opinions → synthesis only
    context: "<optional: 1–3 sentences of relevant conversation/project context, including repo paths worth investigating>",
    members: [               // optional override (2–8 members) when the user wants a custom council,
      { persona: "...", brief: "...", model: "opus" }   // e.g. "a council of security experts";
    ]                        // model ∈ fable | opus | sonnet | haiku (invalid/missing → opus)
  }
})

• Pass the user's question verbatim; put any framing into context instead.
• If the question concerns the current project, say so in context and name the relevant paths — members run in the working directory and will investigate the code before opining (something an external-API council could never do).
• Use mode: "quick" when the user says quick/cheap/fast or the question is small. Use full otherwise — the anonymized cross-critique is where most of the value comes from.

Stage 3 — Chairman synthesis (you, after the workflow returns)

The workflow returns { question, mode, council[], peer_reviews[], aggregate_ranking[] } where each council entry has { id, persona, model, stance, confidence, key_points, answer } and aggregate_ranking gives each member's average peer rank (lower = better). Your final message must contain:

1. The synthesized plan/answer — integrate the strongest ideas across members; attribute inline ("The Skeptic flagged…", "The Researcher found…"). Where members disagreed, say how you resolved it and why. Apply your own judgment — you may overrule the council's ranking, but say so explicitly.
2. Council verdict table — Member (model) | Stance | Avg peer rank.
3. Consensus & dissent — the strongest point of agreement, the sharpest disagreement, and any minority dissent the user should weigh before committing.
4. One closing line noting the council was N independent Claude agents (fresh contexts, distinct personas) run on the user's own subscription.

Integrity: never present council members as ChatGPT, Gemini, or any external vendor — they are Claude agents with personas. If the user wants literally-other-vendor opinions, explain that this skill provides independent Claude perspectives instead, and that true cross-vendor councils require API keys (e.g. karpathy/llm-council or gcpdev/llm-council-skill).

Fallback (Workflow tool unavailable)

Reproduce the stages with parallel Agent tool calls: spawn all members in one message (one Agent call per member, model set per the table, fresh prompts containing persona brief + question + ground rules), then in full mode a second parallel round where each member reviews the anonymized answer packet, then synthesize as above.

Error handling

• Members that fail or return nothing are reported by the workflow; proceed with the responders and tell the user which seats were empty.
• If the whole workflow fails, tell the user and offer your own single-perspective analysis instead.