Council

Council — 5-Model Strategic Advisory

Ask the same question to five LLM seats, each locked to a thinking role, then have the orchestrator (Claude) synthesize. Quality matters more than latency or cost (~$0.50–$1.00 per call, 60–180s wall time).

The five seats

The orchestrator (whichever model is running the skill) does the Blue Hat synthesis pass. Including the orchestrator as a seat would be self-querying.

Each seat is overridable via environment variables — see "Configuration" below.

Preparation protocol

The brief IS the work. Budget 15–30 minutes assembling it before launching any seats.

1. Research the topic. If your stack has a memory/recall layer, query it first.
2. Read the full source material. Read the article, not a summary. Read the code, not a description of it.
3. Pull relevant data — analytics, metrics, anything that changes the question.
4. Steel-man both sides before writing the brief, so the council isn't anchored on yours.

Write the question to docs/council-sessions/YYYY-MM-DD-<slug>/question.md. The file should contain: decision required, why now, full source material, performance data, historical context, the case for, the case against, alternatives, specific numbered questions, constraints.

Ground rules:

• Never send a summary. Seats form opinions from primary sources.
• Always include data. No data = the session is not ready.
• Steel-man both sides to prevent anchoring bias.
• Ask specific questions, not "what do you think?".

Execution

1. Extract the question from the argument or conversation context. If none, ask the user.
2. Create the output directory: docs/council-sessions/YYYY-MM-DD-<slug>/. <slug> is the first 5 words of the question, lowercase, hyphenated, max 50 chars.
3. Write the brief to question.md in that directory.
4. Copy the question to a temp file (/tmp/council-question-<timestamp>.txt) to avoid shell-argument-length limits.
5. Launch all five seats in parallel via the Agent tool. Each agent reads the temp file, invokes its model via the CLI listed in "Provider invocation" below, and writes the full response to the matching file.
6. Verify all five output files exist and have content (>200 chars each). Note any failures.
7. Synthesize. Read all five responses and write synthesis.md.
8. Present the synthesis. Delete the temp file.

Output structure

docs/council-sessions/YYYY-MM-DD-<slug>/
  question.md       — the brief that was sent to every seat
  theorist.md       — Theorist response
  validator.md      — Validator response
  provocateur.md    — Provocateur response
  diagnostician.md  — Diagnostician response
  field-engineer.md — Field Engineer response
  synthesis.md      — consensus + disagreements + recommendation

Synthesis template

# Council Synthesis — <date>

## Question
<one-sentence statement of the decision>

## Consensus
- <points where 4+ seats agree>

## Disagreements
| Topic | Theorist | Validator | Provocateur | Diagnostician | Field Engineer |
| ----- | -------- | --------- | ----------- | ------------- | -------------- |
| ...   |          |           |             |               |                |

## Recommendation
<the orchestrator's synthesis based on all five inputs + own judgment>

## Seat summaries
- Theorist: <one line>
- Validator: <one line>
- Provocateur: <one line>
- Diagnostician: <one line>
- Field Engineer: <one line>

Provider invocation

The skill assumes CLI access to the underlying providers. Use whichever wrapper you prefer — Codex CLI for OpenAI, Antigravity CLI (agy) for Google (replaces the sunset Gemini CLI on consumer plans as of 2026-06-18; enterprise plans may still use the legacy gemini binary), curl for the OpenAI-compatible endpoints (Moonshot, DeepSeek). Model identifiers like gemini-3-pro are unaffected by the CLI rename and resolve under agy directly. When scripting agy, use agy -p "<prompt>" </dev/null — the stdin redirect is mandatory or agy deadlocks on TTY input. Example for Moonshot:

curl -s https://api.moonshot.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $KIMI_API_KEY" \
  -d "$(jq -n --rawfile q /tmp/council-question.txt \
      '{model:"kimi-k2.6",messages:[{role:"user",content:$q}],temperature:1}')" \
  | jq -r '.choices[0].message.content' > theorist.md

The Agent tool should run the curl inside its sandbox and write the result via the Write tool so the file lands in the right session directory.

Configuration

Environment variables (all optional — defaults are picked to maximize each seat):

Codex CLI and Antigravity CLI (agy) auth in their usual config files (ChatGPT subscription auth or Google subscription auth). API keys are only needed for Moonshot and DeepSeek. GEMINI_API_KEY is unchanged — Google did not rename the env var.

Collaborative mode (--collab)

Two rounds. R1 is the parallel single-shot pass above. After R1, the orchestrator reads all five R1 responses and writes a per-seat R2 prompt containing the other seats' positions and specific challenges. R2 runs in parallel like R1. The synthesis then weights R2 positions more heavily than R1 (seats have seen each other's arguments and can update).

Use --collab for high-stakes decisions where the R1 disagreement is large and worth spending another pass on.

When to use

• Strategic decisions: architecture, pricing, positioning
• Format/design choices where the trade-off space is non-obvious
• Risk assessment before committing to a major change
• Tiebreakers between two options that look equivalent
• Diverse-perspective research questions

When not to use

• Tactical questions with a clear right answer (a single capable model is faster)
• Anything where you wouldn't act on the answer anyway
• Pure factual lookups (use a search tool)

Known gotchas

• Kimi thinking ON often hurts Provocateur quality. Locking the model into a reasoning trace before the first token makes it commit to the wrong framing. Default is OFF.
• Steel-manning bias. If you only steel-man the side you already agree with, the council will mirror your bias. The "case against" must be at least as strong as the "case for".
• Empty responses. A seat occasionally returns ~50 chars of "I cannot answer" filler. Check file size before synthesizing; if any seat returns <200 chars, retry that seat before falling back to a 4-seat synthesis with a noted absence.
• Rate limits on the OpenAI-compatible endpoints. Moonshot and DeepSeek throttle hard on parallel calls from the same IP. If both seats fail simultaneously, serialize them.

Anti-patterns

• Sending a summary instead of the source material. The seats then synthesize your summary, not the underlying truth.
• "What do you think?" prompts. The seats default to bland consensus on open-ended prompts. Ask numbered, decision-shaped questions.
• Asking the same question every week. Council output is for decisions, not check-ins.

Validated patterns

• The disagreement matrix in the synthesis is often more useful than the recommendation itself. The places the seats split signal where the real bet is.
• Pasting the entire source document, not a description of it, is the single biggest quality lever.

Self-improvement

This skill ships with a lightweight feedback loop. Adopt or ignore — the skill works without it.

Trigger a review when:

• The user explicitly corrects a recommendation (strongest signal — log immediately; 2–3 corrections on the same theme → promote to body).
• A golden case that previously passed starts failing.
• A novel input category recurs across sessions.
• learnings.md crosses ~100 bullets (consolidation time).
• The skill mis-triggers (loads when it shouldn't, or fails to load when it should — usually means the description: field needs work).
• A council seat's underlying model gets a major version bump.

Consolidation pass (5–10 min, weekly or threshold-driven):

• Each learnings entry gets one fate: apply, capture, or dismiss.
• Apply = merge into Known gotchas / Anti-patterns / Validated patterns.
• Capture = leave in the log for now.
• Dismiss = delete.
• Bump last-consolidated: in frontmatter.