Council

A 5-advisor decision council with anonymous peer review and chairman synthesis. Inspired by Andrej Karpathy's LLM Council concept and Ole Lehmann's implementation. Upgraded with full ecosystem integration.

How to use

• /council <question or decision> — run the full 5-advisor council.
• Trigger phrases (case-insensitive, start of message):
  • council this: <question>
  • war room this: <question>
  • pressure-test this: <question>

When to use

• A strategic decision with significant consequences.
• A go / no-go that needs multi-perspective pressure testing.
• Architecture or product direction choices that need adversarial review.
• Any decision where hidden risks or missed opportunities could be costly.
• Explicit user trigger.

Required input contract

Before running Phase 0, identify:

• The decision — clearly stated, not implied.
• Stakeholder — who will act on the recommendation.
• Decision shape — binary go/no-go, A-vs-B, multi-option, or strategic stance.
• Reversibility — can the decision be undone if wrong?
• Time horizon — Monday-morning action, or longer planning cycle?

If the decision shape or stakeholder is unclear, ask one sharp question before running Phase 0.

Protocol — six phases in order

When triggered, execute ALL phases in order. Never skip any phase.

Phase	Output
0. Intelligence gathering	Intelligence Brief (≤ 300 words) covering Domain, Hard Data, Prior Decisions, Known Context, Data Gaps
1. Context extraction	2-3 sentence decision summary
2. Five advisors	300-600 word response per advisor, each referencing Intelligence Brief
3. Anonymization + peer review	5 reviewers answer Q1 (strongest), Q2 (biggest blind spot), Q3 (collective miss)
4. Chairman synthesis	Executive Summary, Recommendation, Confidence, Next Step, Risks, Opportunities
5. Output	HTML report + Obsidian note (if vault) + Markdown transcript
6. Follow-up	TaskCreate + memory record + report path

Full per-phase procedure: references/council-workflow.md.

The five advisors (decision-critical routing)

Each advisor produces a 300-600 word response that must reference the Intelligence Brief. Generic advice is unacceptable.

#	Advisor	Stance	Signature question
1	Contrarian	Assumes there IS a hidden disaster; digs until found or proven absent	"What will kill this that nobody is talking about?"
2	First Principles Thinker	Strips every assumption; rebuilds from foundations	"What are we actually trying to achieve, and is this even the right question?"
3	Expansionist	Finds every adjacent opportunity, bigger version, multiplier	"What's the 10x version of this, and what are you leaving on the table?"
4	Outsider	Zero prior knowledge; answers purely from what is written	"Reading this cold — what makes no sense, what's unexplained, what would I ask?"
5	Executor	Demands concrete Monday-morning first steps	"What exactly do you DO on Monday morning? If you can't answer that, this is vapor."

The advisor lens is fixed. The domain (selected in Phase 0) determines which data the lens is applied to. Never swap roles or collapse advisors — divergence is the source of the council's value.

Full advisor profiles + per-domain specialization (engineering, trading, strategy, product, etc.): references/roles-and-routing.md.

Decision logic — confidence scoring

Score	Threshold	Conditions
HIGH	> 85%	Strong consensus, hard data supports, low downside if wrong
MEDIUM	50–85%	Split opinions or incomplete data; weight of evidence favors
LOW	< 50%	Genuine uncertainty; recommendation is best-guess; consider follow-up council

Drop one band if: data gaps unresolved · Contrarian's flaw has no mitigation · suspicious unanimous agreement · reviewer Q3 materially shifts the decision.

Full chairman synthesis structure + decision patterns (binary / A-vs-B / multi-option / strategic) + dissent resolution: references/decision-frameworks-and-synthesis.md.

Authority, safety, truthfulness constraints

These are non-negotiable. Skipping any is grounds to rerun the phase.

1. Phase 0 is mandatory. Advisors without data are just guessing.
2. Peer review is mandatory. It catches groupthink — Q3 must always surface at least one collective miss.
3. Anonymization mapping is revealed ONLY at the very end of the transcript.
4. Advisors must diverge. Two advisors with the same conclusion is one wasted advisor — force the lens to apply harder.
5. Every response references specific Intelligence Brief data. No generic advice.
6. Contrarian stays brutal. Their job is to find problems, not be polite.
7. Executor demands concrete Monday-morning steps. "Think about X" is not an answer — the recommendation is incomplete without a named action, tool, person, and deliverable.
8. HTML report MUST be written to disk. Not just displayed inline.
9. Memory record MUST be created in Phase 6. Future councils build on it.
10. Chairman picks a side. No "it depends", no "pros and cons", no "depends on your goals". If the decision is bad, say so; if good, say so.
11. Hermes multi-platform delivery is T2. Ask before sending.

Tone rules

• All advisors and reviewers: brutally honest, zero flattery, zero corporate speak, no hedging.
• Chairman: decisive and neutral. Picks a side.
• Banned: "great question", "pros and cons to both", "ultimately it depends on your goals".

Validation gates

A council run must clear all six gates before delivery.

Gate	Pass criterion
Phase 0	Brief ≤ 300 words; Data Gaps honest
Phase 1	Summary names core decision + stakeholder
Phase 2	Each advisor 300-600 words, references brief, no two converge
Phase 3	All 5 reviewers answered Q1/Q2/Q3; Q3 produced a collective miss
Phase 4	Recommendation specific; confidence rated; next step named
Phase 5	HTML on disk; transcript ends with mapping; Obsidian skipped silently if unreachable

Full anti-pattern catalog + troubleshooting recipes + MCP-outage recovery: references/validation-troubleshooting-and-antipatterns.md.

Output expectations

Three outputs, in this exact order:

1. HTML report — ~/.claude/council-reports/council-<YYYY-MM-DD>-<slug>.html; self-contained inline styles; display inline too.
2. Obsidian note (only if vault accessible) — Council Reports/council-<YYYY-MM-DD>-<slug>.md; skip silently on MCP failure.
3. Markdown transcript — full raw transcript ending with the A-E mapping reveal.

Followed by Phase 6: TaskCreate from Executor's action, memory entity, report path surfaced. Full schemas + slug rules + worked run: references/examples.md.

Reference map

Need	Read
Full per-phase procedure (0-6) with domain table, MCP-source map, anonymization rules	references/council-workflow.md
Five advisor profiles + per-domain specialization (engineering, trading, strategy, etc.)	references/roles-and-routing.md
Chairman synthesis structure, confidence framework, decision patterns, dissent resolution	references/decision-frameworks-and-synthesis.md
Validation gates per phase, anti-patterns with detection + remediation, recovery procedures	references/validation-troubleshooting-and-antipatterns.md
Output schemas (HTML, Obsidian, transcript) + slug rules + complete worked run	references/examples.md