Llm Council

---
name: llm-council
description: "Run a five-advisor council with blind peer review for consequential decisions. Use when the user requests a council or a multi-perspective review of a consequential decision; skip quick critiques and routine factual or writing tasks."
---

# LLM Council

Help the user examine a decision beyond the framing and assumptions in their question. Use five distinct perspectives, blind peer review, and a chair's synthesis. The goal is better judgment, not a vote or guaranteed correctness.

## Prepare

Frame a neutral decision brief from the user's question and relevant available context: objectives, options, constraints, stakes, facts, and unknowns. Gather context proportionately; ask only when a missing detail would materially change the decision and cannot be resolved from available context or an explicit conditional assumption. Keep evidence distinguishable from assumptions.

Read [subagent assignments and dispatch templates](references/subagents.md) for the roles and their inputs. Use the current harness's available subagent tools. When dispatch, isolation, capacity, or role selection needs clarification, read only the matching adapter: [Codex](references/codex.md), [Cursor](references/cursor.md), [Command Code](references/commandcode.md), [Grok Build](references/grok-build.md), [OpenCode](references/opencode.md), [Factory Droid](references/factory-droid.md), or [Claude Code](references/claude.md). Identify the harness from the active environment, not the selected model. Other harnesses require capability verification before use. Reuse capability checks established in the current session unless the runtime, configuration, or dispatch mode changes. Check each reviewer's supplied input packet; record unobservable settings as unverified and do not claim unestablished isolation.

## Run the council

1. **Independent perspectives.** Spawn five advisors, one per lens: Contrarian, First Principles Thinker, Expansionist, Outsider, and Executor. Give each the same brief and its assigned lens, without other answers or a preferred conclusion. Run concurrently where supported; isolated batches are acceptable.
2. **Blind peer review.** After all answers arrive, remove explicit author/role labels and shuffle them into Response A–E. Retain the mapping for the chair and transcript. Preserve answer wording and meaningful formatting; remove only explicit identity labels and record those removals. Give five fresh reviewers the same brief and complete anonymized answers, without the mapping or other reviews. Keep letter identities fixed and rotate presentation order: ABCDE, BCDEA, CDEAB, DEABC, EABCD. Record each reviewer’s order; this balances positions, not identities or factual knowledge. Each identifies the strongest argument, the biggest blind spot, and what all five missed.
3. **Synthesis.** Give a separate chair the original question, brief, named answers, mapping, and reviews. Ask for agreement, disagreement, insights from review, a clear recommendation, and one concrete next step. The chair may favor a well-supported minority argument.

## Preserve what makes this useful

- Assemble each dispatch with the role template and check its required inputs before sending. Retain a dispatch record and complete returned text for the transcript; never substitute coordinator summaries for worker answers. Use the recording rules in the shared assignments to distinguish verified capture, coordinator copies, and reconstructions. Shared record components may be stored once; workers still receive complete packets.
- Keep mappings, prior reviews, and other council records out of worker-accessible context where supported. Read-only access does not prevent reading earlier files. Keep working records outside worker access when possible and write public artifacts after the chair finishes; delaying writes alone does not establish isolation. Disclose remaining shared-file or ambient-instruction access.
- Separate contexts matter more than simultaneous starts. Reviewers must not inherit an advisor's identity or the coordinator's preferred answer.
- Ask for direct, lens-specific analysis without invented evidence or forced disagreement. Shared-model agreement is not factual verification; anonymous labels cannot conceal every stylistic cue.
- Use available evidence. When a consequential factual claim needs checking, verify it through appropriate available tools or identify it as unresolved. Council discussion alone does not validate it.
- Correct material input defects using the shared assignments’ replacement rules for any role; preserve superseded attempts and affected downstream dependencies.
- Recover outstanding results through the supported lifecycle before declaring a round incomplete; report a blocker when recovery is unavailable. If delegation or isolation is unavailable, or a round cannot finish, report the limitation. Do not pass off one-agent roleplay or missing reviews as a completed council.

## Deliver

Save `council-report-{timestamp}.html` and `council-transcript-{timestamp}.md` in the requested directory, defaulting to the current project. Use matching UTC timestamps such as `20260927T120000Z` and avoid overwriting past sessions. If recovery is impossible, save both artifacts with a prominent incomplete status, the returned material, and the blocker; do not invent a chair verdict or present the council as complete.

The HTML report should make the question and verdict easy to scan, show actual agreement/disagreement, and offer collapsible advisor responses and review highlights. Keep it self-contained, readable, and usable at phone width: wrap long text and wrap or horizontally scroll wide tables. Escape inserted text. Preview it at desktop and phone widths when supported.

The Markdown transcript records the original question, brief and sources, all returned answers and reviews, the revealed mapping, and full verdict. Include dispatch records with their capture status and reviewer presentation orders, plus actual execution details and material gaps: completed rounds, batching, isolation limits, and models if known. Record returned analyses, not hidden internal reasoning.

Before delivery, check decision-critical arithmetic, dates, units, and claims of agreement against the supplied evidence and returned analyses, using deterministic tools where available. Keep the check proportionate to what could change the decision and respect the run’s research boundaries. Preserve original worker text; put discovered errors and unresolved checks in a separate, clearly labeled qualification beside the recommendation. Do not present a recommendation undermined by those checks as ready to act on.

Return the recommendation, or the blocker for an incomplete council, and links to both files. Check that the artifacts faithfully contain the returned council output and accurately state completion.