llm-council-skills

Designs a human AI governance council - hybrid management council plus board escalation, charter and conflict-of-interest templates, decision-rights matrix, incident escalation, KPIs, and US/EU/UK/China compliance lanes. Use for "set up an AI governance board/council", "AI review committee charter", "model launch approval process". Not for multi-model LLM ensembles or technical safety evals.

Combines multiple LLM answers or judge verdicts robustly - correlated-error limits of majority voting, reliability weighting, side-swapped pairwise comparison, judge-synthesizer separation, and surfacing dissent. Use for "combine/aggregate model outputs", "LLM judge panel/jury", "majority vote across models", "rank council responses". Not for pipeline topology, member choice, or stage prompts.

Designs LLM council orchestration - parallel fan-out, anonymized peer review, chairman synthesis, topology choice by task shape, debate round caps, stopping rules, graceful degradation. Use for "build/design an LLM council", "multi-model deliberation architecture", "structure a debate/review/synthesis pipeline". Not for member selection, prompt wording, aggregation math, or cost tuning.

Cuts cost and latency of multi-model LLM pipelines - route-first gating so the council fires only on hard queries, calibrated-confidence escalation, prompt caching, batching, tail-latency budgeting. Use for "the council/ensemble is too expensive or slow", "route to a cheaper model then escalate", "reduce multi-LLM cost". Not for deciding whether to adopt a council or aggregation choice.

Defends LLM councils against failure modes - sycophancy and groupthink (structured dissent, round caps, no early-stop-on-consensus), correlated-error monitoring, and prompt-injection amplification containment. Use for "models keep agreeing", "council groupthink", "multi-agent prompt injection", "consensus feels fake". Not for single-agent jailbreak hardening or member selection.

Implements an LLM council over headless coding harnesses in plain Python - one router over Claude Code, Codex CLI, and API models, subprocess hygiene, JSON contracts with tolerant extraction, per-member failure isolation, boss-synthesis fallback. Use for "council of Claude Code/Codex CLIs", "run coding agents headless as council members". Not for API-only orchestration frameworks.

Selects and role-designs LLM council members - provider-diverse 3-5 model panels, cognitive lenses over personas, role-to-model routing, anonymization and identity effects, persona-drift countermeasures. Use for "which models should sit on the council", "pick council members", "design advisor personas/lenses". Not for pipeline topology, aggregation math, or stage prompt wording.

Writes the prompts for each LLM council stage - forced-perspective advisor prompts, anonymized peer review with a parseable ranking contract, decisive chairman synthesis that may side with a dissenter, rebuttal rounds. Use for "write the council/advisor/chairman prompts", "peer review prompt", "synthesis prompt". Not for pipeline architecture, member selection, or aggregation rules.

Decides whether an LLM council (multi-model deliberation/ensemble) beats a single strong model - task-type triage, compute-normalized evidence, cost-per-resolved-outcome, benchmarking both paths on your workload. Use for "should I use multiple models", "is an LLM council/ensemble/debate worth it", "single model vs council". Not for designing the pipeline, picking members, or cutting cost.