Skip to content

/council

Convene a multi-perspective 'council' on an open question

Category: ideation
Field: general
License: MIT
Updated: 2026-04
Stages: rq-formulation · hypothesis-generation

/council — Parallel Critics + Separate Synthesis

v1.3 — Removed the Phase 4 "rate-limit advisory" line: it depended on response headers that aren't accessible under OAuth-only auth, so it printed "unavailable" on every run. The "Ctrl-C to abort" affordance was also phantom — print and dispatch happen in the same model turn, so there's no actual abort window. Replaced with a clean dispatch announcement: $PANEL_NAME panel: dispatched N critics on $COUNCIL_CRITIC_MODEL. v1.2 added --chef-skill. v1.1 added --mixed (cross-vendor peer). v1.0: dispatches N critic agents in parallel, runs a separate synthesis pass; hard cap 5; single-round only; never majority-votes on narrative output.

Dispatches N critic agents in parallel, collects raw outputs, and runs a separate synthesis pass. Hard cap of 5 critics. Single-round only. Never majority-votes on narrative output.

Quick Start

The fastest path that works out of the box:

Text Only
/council --chef-skill <topic or file:path>

This uses a hardcoded 3-role panel (skill-engineer + UX-for-tools + non-expert-adoption) implemented as general-purpose subagents with inline role-strings. No persona files required — runs immediately.

For broader use (plan / paper / decision / grant panels), see "Default panels" below — those require persona agent files in ~/.claude/agents/.

Config

Text Only
COUNCIL_CRITIC_MODEL=opus       # or sonnet for cheaper runs
COUNCIL_SYNTHESIS_MODEL=opus
COUNCIL_MAX_CRITICS=5
COUNCIL_DEFAULT_N=3

Models are set here, not in persona frontmatter. Next-model migration = edit two lines.

Invocation

Primary: /council <topic> or /council file:<path> [flags]

Natural-language triggers: - "spin up a council on X" / "council this" / "panel of advisors" - Chef-skill voice triggers → auto-set --chef-skill: "council this skill" / "skill council" / "council on this tool / command / workflow" / "bespoke skill council" - Meta-reference guard: sentences about councils as a topic are not invocations. On low confidence, ask one line.

Flags

Flag Effect
--panel a,b,c Explicit persona list (names without -agent suffix)
--n K Number of critics (default 3, max 5)
--type plan\|paper\|decision\|grant Forces panel default + synthesis branch
--mixed [codex\|gemini] Swap ONE Claude critic for a cross-vendor peer
--chef-skill Hardcoded 3-role skill/tool-design panel. Bypasses panel resolution and persona-file check. Works out of box.

Default panels (resolve from --type or keyword inference)

These require persona agent files in ~/.claude/agents/. The file naming convention is <persona-name>-agent.md or experts/<persona-name>.md.

Task type Default panel N
Plan / architecture review skills-engineer, skeptic, pre-mortem, budget-hawk 4
Paper / proposal review academic-editor, harsh-referee, methodologist 3
Grant proposal academic-editor, harsh-referee, methodologist, grant-strategist, funder-officer 5
Decision support skeptic, pre-mortem, chief-of-staff 3
Ambiguous prompt user to pick —

If you don't have these persona files yet, use --chef-skill for skill-design topics or run with explicit --panel listing personas you do have. The skill aborts cleanly if a referenced persona file is missing — no silent fallback to general-purpose.

Keyword inference (used if --type not given): "plan / architecture / design" → plan; "paper / referee / submission / manuscript" → paper; "proposal / grant / funder / NSF / Gates" → grant; "should I / go-no-go / accept / reject / decide" → decision; else ambiguous → ask.

Dispatch phases

Phase 1 — Parse and resolve

  1. Parse topic + flags.
  2. If file:<path>, Read the file; that is the content for the critics.
  3. Resolve panel:
  4. Explicit --panel → use verbatim
  5. --type X → use that row's panel
  6. Keyword inference → pick type
  7. Ambiguous → ask one line: "Panel type: plan / paper / decision / grant?"
  8. Cap N at COUNCIL_MAX_CRITICS. If user asks for >5, refuse with: "Hard cap is 5. Pick a tighter panel."

Phase 1.5 — --chef-skill short-circuit

If --chef-skill flag is set, skip Phases 2 (meta-guard) and 3 (persona-file check). This branch ships a hardcoded 3-role panel for skill/tool-design review, dispatched as general-purpose subagents with inline role-string prefixes — not persona files.

Pre-populate these 3 role-strings and jump directly to Phase 5 dispatch:

  1. Skills engineer — "Will this skill work, will it last, and does the abstraction earn its keep? Focus on: invocation discoverability, prompt-budget discipline, failure recovery, and whether it duplicates existing skill/rule coverage."
  2. UX-for-tools critic — "Where does this add friction, cognitive load, or break on voice input? Focus on: flag syntax vs. natural-language triggers, multi-step workflow friction, CLI ergonomics."
  3. Non-expert adoption critic — "Will a junior user with <10h Claude Code successfully invoke and recover from failure? Focus on: discoverability of invocation, clarity of error messages, cold-start onboarding."

Panel type for synthesis: plan. N critics: 3.

Phase 2 — Meta-reference guard

If invocation is via natural-language trigger (not explicit /council), classify whether this is an invocation or a meta-reference. Dispatch one short Task call to classify; if META or UNCLEAR, ask one line before dispatching critics. Skip this phase if --chef-skill is set.

Phase 3 — Persona file existence check

For each persona in the resolved panel, Glob both: - ~/.claude/agents/{name}-agent.md - ~/.claude/agents/experts/{name}.md

If any persona file is missing, abort with a clear error listing the missing personas. No silent fallback to general-purpose. Skip this phase if --chef-skill is set.

Phase 4 — Dispatch announcement

Print one line before the parallel Task calls:

Text Only
$PANEL_NAME panel: dispatched N critics on $COUNCIL_CRITIC_MODEL

No terminal period (this is a status line, not prose). $PANEL_NAME is one of plan / paper / decision / grant / chef-skill. If $COUNCIL_CRITIC_MODEL resolves to a long versioned string, display the short label (e.g., opus-4.7).

Phase 4.5 — Peer swap (only if --mixed flag set)

If --mixed codex|gemini was passed, swap ONE Claude critic for a cross-vendor peer. Preserves the 5-critic cap.

Persona-to-vendor auto-mapping:

Peer vendor Best swap target Rationale
Codex skeptic, pre-mortem, or harsh-referee Codex strong on empirical rigor + failure modes
Gemini completeness-checker, domain-expert, or academic-editor Gemini strong on long-context synthesis + broad coverage

Pick the first match from the resolved panel. If none match, swap the last critic (deterministic).

The swapped Claude critic becomes a peer-vendor critic via Bash dispatch (Codex CLI or Gemini CLI). The other N-1 critics stay as Task dispatches. Direction: Claude → peer only. Do NOT dispatch peer-first-then-Claude (asymmetric dismissal).

If the peer binary is missing, print one line ("Gemini CLI not installed; falling back to all-Claude council") and continue with the unswapped panel. Graceful degradation.

Phase 5 — Parallel Task dispatch

Send one message with N Task tool calls (true parallelism). Each call: - subagent_type: persona slug (e.g., skeptic-agent) - model: value of COUNCIL_CRITIC_MODEL (explicit, overrides any persona frontmatter) - description: 3-5 word summary - prompt: the topic or file content + brief instruction to produce raw critique in the persona's output shape

Mixed mode addition: If Phase 4.5 swapped a critic, include ONE Bash call in the same parallel message that invokes the peer binary. Output goes to /tmp/council-peer-$RUN_ID.md. The synthesis pass in Phase 7 reads the peer output alongside the Claude critic outputs.

--chef-skill mode: Dispatch 3 general-purpose Task calls in parallel (not persona-agent slugs). Each call's prompt opens with the corresponding role-string from Phase 1.5 as a prefix, followed by the topic/file content and the instruction "Produce raw critique in this role's voice. End with VERDICT: APPROVE | REVISE + one-line rationale."

Phase 6 — Collect

Collect raw critic outputs. Do NOT inline-synthesize.

Phase 7 — Separate synthesis dispatch

One more Task call: - subagent_type: general-purpose - model: COUNCIL_SYNTHESIS_MODEL - prompt: synthesis template + the panel-type branch + raw critic outputs + original content

Panel-type branches: - plan → Verdict (ship/revise/kill) → Top 3 blockers → Top 3 patches → per-critic verdict - paper → Referee recommendation → Contribution-claim strength → Required revisions → Optional revisions → per-critic verdict - decision → Recommended action → Top 3 risks → Top 3 reasons in favor → per-critic verdict - grant → Paper branch + funder-officer fundability verdict

Phase 8 — Emit

Emit the synthesis. Include raw critic outputs inside a <details> collapsible at the bottom for verification.

Phase 9 — Log

Append one row to a council invocation log (e.g., ~/.claude-assistant/logs/council-invocations.csv):

Text Only
timestamp,run_id,topic,panel_resolved,critic_count,synthesis_verdict,wall_time_sec,model_critic,model_synthesis

Why a Council

Multiple critics with explicit lenses outperform one general critic on review/decision tasks. The pattern:

  • Parallel — each critic sees the same input and produces output independently
  • Separate synthesis — a fresh-context model reads the raw critic outputs as data, not as conversation history
  • Single round — no iterative debate. Round 2+ documented to drift toward conformity (see 2025-2026 multi-agent literature)
  • Hard cap 5 — beyond 5 critics, you get diminishing returns and harder synthesis

Out of scope (v1.0+)

  • Round 2 / iterative debate (single-round only)
  • Majority voting on narrative output (conformity bias)
  • LLM-proposed persona generation (--chef mode without -skill): requires a router. Hardcoded --chef-skill is the minimum-viable path for now.

Performance Logging

Bash
echo "$(date +%Y-%m-%d),council,TOOL_CALLS,${N_CRITICS}-critics-${PANEL_TYPE}" >> ~/.claude-assistant/logs/skill-performance.csv

Setup Notes

  • For chef-skill mode: nothing required. Works out of box.
  • For default panels: place persona agent files in ~/.claude/agents/. The site provides one starter agent (agents/proposal-critic-agent.md); add more as you build out your panel library.
  • For mixed mode: install the peer CLI (Codex via macOS app, Gemini via npm install -g @google/gemini-cli). The skill degrades gracefully if either is missing.