| name | council |
| description | Run a SplitBrain.AI LLM council review for high-impact decisions. Use when evaluating architecture tradeoffs, orchestration design, schema changes, retrieval/search behavior, agent governance, or long-term infrastructure decisions. |
| argument-hint | Decision topic and scope to review |
| user-invocable | true |
| disable-model-invocation | false |
LLM Council Review for SplitBrain.AI
Inherits from task-core-loop.
Use this skill to run a structured, evidence-first, multi-perspective decision review.
Subagent Permission Policy
- Default to a single-agent, self-simulated council analysis.
- Do not invoke subagents for council seats unless the user gives explicit permission in the current request.
- If explicit permission is not present, proceed with self-simulated seats and document that choice in the report.
Outcome
Produce a council report that includes:
- A one-sentence decision statement
- Seat-by-seat recommendations with confidence percentages
- Explicit disagreement and dissent
- Risks, assumptions, open questions
- Acceptance criteria and validation gates before implementation
Use When
Use this workflow when decisions affect:
- Orchestration architecture or node client dispatch strategy
- Memory record shape or schema fields in the wiki service
- Retrieval quality, ranking behavior, or search configuration
- Chat or agent write behavior and evidence traceability
- Dashboard or MCP interface conventions
- Long-lived architecture or migration cost
- Model provider selection or inference pipeline design
Do not use this workflow for quick lookups, one-off formatting, or routine implementation decisions.
Inputs
Collect these inputs before running the workflow:
- Decision topic and one-sentence decision question
- Scope of impact: orchestration, wiki/schema, retrieval/search, chat/agent, dashboard, or mixed
- Primary evidence from
Docs/Memories/, Docs/Plans/, and source code
- Any known stale docs, assumptions, or unresolved constraints
Procedure
-
Classify the decision.
If this is low impact (cleanup, formatting, direct retrieval), skip council and do a normal edit or lookup.
If this is high impact, continue.
-
Build a bounded evidence pack.
Prefer MCP wiki tools and local project evidence first.
Minimum pack should include:
- Relevant memories in
Docs/Memories/
- Relevant plans in
Docs/Plans/
- Source-linked code evidence when claims depend on implementation
- Tests and benchmarks when behavior is affected
- If tests or benchmarks are not possible in exceptional circumstances, document why, add alternative evidence, and define a follow-up validation gate
- Select council seats.
Use at least 3 seats (default) and expand toward all 6 for major architecture decisions:
- Source-Grounded Archivist
- Data Model Architect
- Retrieval Specialist
- Human Learning Advocate
- Skeptical Reviewer
- Synthesizer
- Run independent seat reviews.
Give each seat the same evidence and require:
- Findings
- Risks
- Recommendations
- Assumptions
- Open questions
- Confidence percentage
Subagent usage note: only run seats via subagents when explicit user permission is provided. Otherwise run all seats in-process.
-
Branch on disagreement.
If seats materially disagree, do not flatten to consensus.
Record the disagreement and identify missing evidence that would change the outcome.
-
Synthesize a decision.
The Synthesizer must separate:
- What changes now
- What is deferred
- What evidence gates must be passed before implementation
-
Record the result.
Write or update the decision report with dissent visible, confidence values, and validation criteria.
-
Gate implementation.
Only proceed to code changes or broad migrations after acceptance criteria are explicit and verifiable.
Decision Branches
-
Branch A: Convention-first vs schema-first
If the concept is immature or primarily documentation/process, prefer convention-first plus validation probes.
Promote to schema only after patterns are stable and machine parsing or enforcement is clearly needed.
-
Branch B: Retrieval/search sensitive changes
If ranking or recall quality may change, prefer measurable retrieval checks (tests/benchmarks) and explicit rollback notes.
If measurements are not feasible in an exceptional case, require documented rationale plus a time-boxed follow-up measurement plan.
-
Branch C: Chat/agent write behavior
If trust or safety boundaries are affected, require evidence traceability and approval-gated write behavior.
-
Branch D: Evidence weakness
If evidence is thin or stale, defer implementation and define what evidence must be gathered.
Completion Checks
A council review is complete only when all are true:
- Decision statement is explicit
- Evidence links are listed and source-grounded
- Each seat includes confidence and blocking concerns
- Dissent is visible, not merged away
- Acceptance criteria are testable or reviewable
- If tests/benchmarks are omitted, exception rationale and follow-up validation gates are explicit
- Open questions are documented with next-step owners or gates
Report Template
Use this structure in the final output:
# Council Review: <Decision>
## Decision
<one sentence>
## Evidence Reviewed
- <memories/plans/source links>
- <tests/validation commands>
## Findings
| Seat | Recommendation | Confidence | Blocking concern |
|---|---|---:|---|
| Source-Grounded Archivist | ... | 0.85 | ... |
| Data Model Architect | ... | 0.80 | ... |
| Retrieval Specialist | ... | 0.78 | ... |
| Human Learning Advocate | ... | 0.82 | ... |
| Skeptical Reviewer | ... | 0.74 | ... |
## Synthesis
<what changes now vs later>
## Dissent
<unresolved disagreement>
## Acceptance Criteria
- <gate 1>
- <gate 2>
## Open Questions
- <question 1>
- <question 2>
Prompt Pattern
Use this prompt form for each seat:
You are the <seat name> council seat for SplitBrain.AI.
Review <topic> using the supplied wiki and code evidence.
Do not implement code.
Return findings, risks, recommendations, assumptions, open questions, and confidence percent.
Prefer source-linked project wiki evidence over generic advice.
References
Docs/Memories/ — current distilled knowledge
Docs/Plans/MasterPlanV4.md — canonical architecture blueprint
- Source code under
src/Orchestrator.*/, src/SplitBrain.Dashboard/, src/NodeClient.*/