council

Council Skill

Vendored from danielmiessler/LifeOS install/skills/Council (upstream v1.1.20). De-Clauded for this machine: no voice-notification curl, no ~/.claude/ customization path, no execution-log JSONL, no RedTeam cross-references. See Running the members below — the parallel-subagent path is optional here.

What It Does

Runs a multi-agent debate. Custom-composed agents discuss a topic over rounds, respond to each other's points, and expose weak arguments through substantive disagreement. You get a visible round-by-round transcript plus a synthesis. DEBATE runs three rounds; QUICK runs one for a fast perspective check.

The Problem

When you ask one model for an opinion, you get one frame and one set of blind spots. Asking for "pros and cons" gives you a flat list with no one pushing back. Deliberation needs distinct experts who disagree on the merits, so weak arguments surface before you commit. Generic built-in agents tend to produce bland agreement; this skill composes topic-specific agents with conflicting positions.

How It Works

Members discuss the topic in rounds and respond to specific claims from earlier rounds.

Members Are Custom Briefs

Write each council member inline as a short brief — a name, a role, a stance, and what they'll push on. A generic persona is topic-ignorant and produces bland agreement. The friction comes from four different briefs, each with real domain expertise and a distinct analytical angle.

See CouncilMembers.md for the slot guidance and an example brief.

Running the members

Upstream assumes a Task/Agent tool with subagent_type. Pi has no such built-in, so pick whichever of these the current harness supports:

In-context loses true independence — you know what the other members will say. Compensate by committing to each brief's stance hard, and by writing the weakest member's position first so it doesn't get retro-fitted to the conclusion. The transcript is still the deliverable either way.

Workflow Routing

Route to the appropriate workflow based on the request.

Council is collaborative-adversarial: members debate to find the best path. Pure adversarial attack on a single idea is out of scope — say so rather than bending a debate into a teardown.

Quick Reference

Context Files

Core Philosophy

Origin: Council compares informed positions through direct challenges. Domain-specific members respond to each other's claims instead of listing independent opinions.

Agents: Every council member is a custom brief you write for the topic. This gives each member a distinct role, stance, and domain expertise. Generic agents produce generic debate; topic-specific briefs produce sharp, informed debate.

Speed: With parallel subagents, execution is parallel within rounds and sequential between them — a 3-round debate of 4 members is 12 agent calls but only 3 sequential waits (40-90 seconds). In-context, it's one pass and correspondingly slower to read but cheaper to run.

Examples

"Council: Should we use WebSockets or SSE?"
-> Write 4 member briefs (real-time architect, frontend-DX, ops skeptic, analyst)
-> DEBATE workflow -> 3-round transcript

"Quick council check: Is this API design reasonable?"
-> Write 4 member briefs with API-relevant roles
-> QUICK workflow -> Fast perspectives

"Council: Is AI overhyped?"
-> Write briefs: AI builder, security skeptic, pragmatic engineer, evidence analyst
-> DEBATE workflow -> 3-round transcript

Integration

Works well with:

• brainstorm - Council to pick between approaches, then brainstorm the winner into a spec
• mu - When you want the members as long-lived tmux panes you can interrogate

Practices

1. Use QUICK for sanity checks, DEBATE for important decisions
2. Write each member's brief around the specific topic, not a generic role
3. Give each member a distinct stance — four identical agents produce no friction

Gotchas

• Council members are inline briefs. There is no composition tool. Write four different topic-specific briefs; a bare persona-less agent produces bland agreement.
• Debates need substantive disagreement. If all members agree, the topic may not warrant Council.
• More agents ≠ better debate. 4-6 well-briefed members outperform 12 generic ones.
• In-context mode is self-debate. You are simulating disagreement, not sampling it. Worth doing, worth not overtrusting — a convergence you reached alone is weaker evidence than four independent agents landing in the same place.