By name: council description: Convene a panel of 4 Claude subagents (architect, reviewer, simplifier, adversary) on a single hard decision — architecture choice, security review, "should we?" questions. Each gives an independent ≤300-word opinion in parallel; main thread synthesizes consensus and dissent. Result is written to .claude/council/ for later audit. Use only for high-stakes one-shot decisions, not daily code edits. model: opus
When to use
A "stop and think" tool. Invoke for:
- Architecture choices ("Postgres vs SQLite for this service?", "monorepo vs polyrepo?")
- Security/risk reviews ("is this auth flow safe?", "should we expose this endpoint?")
- Reversibility-sensitive decisions ("should we rewrite the X module?", "drop support for Y?")
- Pre-PR sanity check on a non-trivial design
Do not invoke for: daily code edits, single-file refactors, "fix this bug", or anything where the answer is obvious. The council costs ~30–60s and 4× the tokens of a normal turn — only worth it when a wrong call is expensive.
Inputs the user must provide
- The question — a one-paragraph framing of the decision and the leading proposal.
- (Optional) Pointers to relevant files/docs the panellists should ground their opinions in.
If the question is vague ("what should I do about X?"), ask for the leading proposal first. The council critiques a concrete option better than it generates one from scratch.
Steps
1. Build the shared brief (≤500 words)
Compose a single brief that every panellist receives identically. Include:
- The exact question and leading proposal.
- Relevant excerpts from
.claude/CLAUDE.md(≤30 lines) for project context. - Any file paths the user pointed at.
- Constraints the user mentioned (deadline, team size, budget, etc.).
Do not include your own opinion in the brief — that biases the panel.
2. Spawn 4 panellists in parallel
Send a single message with four Agent tool calls in one block so they run concurrently. Each gets the same brief plus a different persona prompt:
| Persona | subagent_type | model | Persona prompt (added to brief) |
|---|---|---|---|
| Architect | general-purpose | opus | "Prioritise long-term maintainability and clean boundaries. What design holds up at 10× the current scale? Recommend the choice that future maintainers will thank us for." |
| Reviewer | general-purpose | sonnet | "Find what could break, get exploited, or fail silently. Assume adversarial users and partial outages. What's the worst plausible failure mode of the leading proposal?" |
| Simplifier | general-purpose | sonnet | "Argue for the smallest possible solution. What's the boring option? What can we not build and still meet the requirement? Push back on any complexity that isn't load-bearing." |
| Adversary | general-purpose | opus | "Argue against the leading proposal. Steelman the opposite choice. If the user is wrong here, why? What are they not seeing?" |
Each panellist's prompt ends with:
Reply in ≤300 words with exactly: Recommendation: (one sentence). Top risk: (one sentence). Confidence: (1–5). Reasoning: (≤6 bullets).
3. Synthesize in main context (do not delegate this)
Once all four return, write the synthesis yourself in the main conversation:
# Council on: <one-line question> (<YYYY-MM-DD>)
## Panel
- Architect (opus): <recommendation> — confidence <n>/5
- Reviewer (sonnet): <recommendation> — confidence <n>/5
- Simplifier (sonnet): <recommendation> — confidence <n>/5
- Adversary (opus): <recommendation> — confidence <n>/5
## Points of agreement
- <bullet>
- <bullet>
## Points of dissent
- <issue>: <who said what, briefly>
## Top risks raised
1. <risk> (raised by <persona>)
2. ...
## Chairman's recommendation
<your synthesis: which proposal, what to mitigate, what to defer. Cite which panellist's reasoning swung the call.>
## Open questions for the user
- <thing the panel could not resolve without more info>
4. Persist the result
Write the synthesis to ./.claude/council/<YYYY-MM-DD>_<slug>.md (create the directory if missing). Slug = first 40 chars of the question, lowercased, non-alphanumerics → -.
This makes past councils greppable later — useful when a similar question comes up.
5. Report to the user
Print the chairman's recommendation + the file path of the persisted council. Do not dump all four panellist replies into the main conversation — they're long and the synthesis already captured what matters. The user can open the persisted file if they want the raw transcripts (include those in the persisted file, after the synthesis).
Hard rules
- Spawn all four panellists in a single message. Sequential spawning quadruples latency for no benefit.
- The synthesis is yours, not a 5th agent. Delegating synthesis loses the user's intent and the conversation context the panel didn't see.
- No leaking your opinion into the brief. The panel is worth nothing if you've already steered it.
- 300-word cap is real. If a panellist runs long, ask them to compress — the synthesis can't hold four 1000-word essays.
- Always persist. A council whose output isn't saved becomes hearsay in 10 turns.
Anti-patterns
- Do not invoke the council for "what's the best way to write this function?" — overkill.
- Do not include the panellist replies in the main reply to the user — only the synthesis.
- Do not let the Adversary persona devolve into nitpicking — its job is steelman, not bikeshed.
- Do not run councils in a loop. If the first one is inconclusive, the answer is "go gather more data," not "ask 4 more agents."
Comparison with external tools
The third-party agent-council (yogirk/agent-council) does the same shape but spawns cross-vendor CLIs (Codex, Gemini) for true model diversity. This skill trades vendor diversity for zero setup — all four panellists are Claude. Use the external tool when the decision genuinely benefits from non-Claude perspectives (e.g. choosing between AI providers themselves); use this skill for everything else.