phrazzld/council

/council

You are the chair, not a participant. Convene a small bench of distinct model families — each in a different generative persona — to think divergently about one question, then synthesize. The point is decorrelated thinking the single orchestrator can't produce alone.

This is the generative sibling of /roster. Roster's adversarial bench reviews an artifact to find the bug (one lens each, artifact-only). A council generates: options, framings, first-principles takes, the non-obvious move, the real disagreement. Reuse roster's dispatch mechanics and model index; this skill owns composition + synthesis.

When to convene (and when not)

• Convene for: a contested design/direction call, a wide-open "what should we build / how should we approach this", a divergence pass before converging on a plan, a brainstorm where one model's taste shouldn't decide. The divergence-for-design-decisions mandate (≥4 structurally distinct voices on non-trivial architecture) is exactly this.
• Don't convene for: a factual lookup, a settled call, anything a native subagent or a moment's thought answers. A council costs real money (each lane is a paid inference run) and your own synthesis time. Options, not obligation.

Compose the bench (the judgment)

Two axes, both must vary — decorrelation comes from family × lens:

1. Distinct model families. 4–6 members, each a different family (Kimi/Moonshot, DeepSeek, Qwen, GLM/Zhipu, MiniMax, …). Same-family variants don't decorrelate — a wide bench of one family is waste. Optionally add a frontier closed model on its own CLI (codex, grok, agy) for extra family spread.
   • Slugs rot in days — never hardcode them. Pull current top models live: roster's references/model-provider-harness-index.md, or the OpenRouter MCP (models-list, benchmarks, model-endpoints for current quality + pricing). pi --provider openrouter --list-models <family> lists live slugs.
2. Distinct generative personas. One per member, pulling in different directions (builder vs simplifier vs user-advocate vs contrarian …). Library: references/personas.md. Compose a bespoke lens for the real question rather than forcing a stock role.

Scale to stakes: a quick divergence is 3 members; a load-bearing design call is 5–6 plus a native subagent or two on further lenses.

Run it

scripts/council.sh fans the bench out in parallel, caps each lane, and collects every output — failed lanes reported, not hidden.

# members.tsv:  label <TAB> cli <TAB> model <TAB> persona   ('#' comments ok)
scripts/council.sh --task /tmp/q.txt --members /tmp/members.tsv --outdir /tmp/council

Write the shared task to one file (inline ALL context — lanes run cold, no shared history). Default cli is opencode (opencode run --model openrouter/<slug>); pi is the lighter no-tools alternative. See /roster for the exact headless forms.

Timeout judgment is part of composition. The script default is 1200s per lane because strong reasoning models often need real wall time for cold, tool-using architecture questions. Use shorter caps only for smoke checks, slug probes, or cheap quick-divergence runs. For load-bearing design calls, keep the default or raise --timeout to 1800s; if a lane hits the cap, report it as caller-capped and rerun or exclude it before drawing model-quality conclusions.

Synthesize as chair (the other half of the judgment)

Reading the lanes is the work — don't just paste them.

• Surface the non-obvious. The value is the idea or framing you wouldn't have reached alone, not the consensus.
• Name the real disagreement. Where lanes genuinely diverge is signal — present the live tension and your call, not an average.
• Don't vote/tally. N models agreeing is weak evidence (shared training); one model's sharp dissent can be the right answer. Weigh, don't count.
• Own the result. Council output is evidence; you decide and are accountable. A rambling or failed lane is a result too — say so.

Gotchas

• Monoculture = theater. Same family across lanes, or all lanes on the same lens, produces correlated mush. Vary both axes or don't convene.
• Cold lanes. Members share nothing but the task file. Inline constraints, goal, and what "good" looks like into the task — a member missing context invents it.
• Stale slugs. A lane failing instantly is usually a dead/renamed slug or an auth lapse, not a verdict — re-check live (roster index / OpenRouter MCP) and re-run that lane.
• Brainstorm ≠ converge in one shot. Use the council to widen; then you (or a focused follow-up) narrow. Don't ask the bench to also pick the winner — that's the chair's job.
• Timeouts are caller evidence. A reasoning lane that hits --timeout was capped by the chair. Use short caps only intentionally, and don't blame the model for a cap chosen too low for the task.
• Cost is real. Bound the bench to the stakes; --timeout is still the runaway guard. The OpenRouter MCP credits-get shows remaining balance.

Composes with

• /roster — dispatch mechanics, live model index, the adversarial-critique counterpart (use that to review, this to generate).
• nous-creative-ideation — a routed library of named ideation methods. Seed a member's persona with a specific method (OuLiPo, TRIZ, lateral provocations), or run the question through one first when the bench risks converging on the obvious.
• The OpenRouter MCP (user-scoped) — live model catalog, benchmarks, pricing, and balance for choosing the bench.