specialist-council

name

specialist-council

description

AI Communication Intelligence layer — the bidirectional cross-specialist consultation protocol that makes specialist agents TALK TO EACH OTHER before an artifact is finalized. Use when a task spans domains and the output needs input from multiple specialists (e.g. marketing copy that must satisfy SEO + GEO + E-E-A-T + brand voice, or a video that needs Remotion + HeyGen + design + brand). Any orchestrator or /spm invokes it to pull structured input from the RIGHT specialists and fold it back in. Triggers on "AI communication intelligence", "agents talking to each other", "cross-specialist", "specialists should consult each other", "get input from all relevant specialists", "council". NOT a builder and NOT a replacement for the single-domain orchestrators — it is the round-trip consultation they lack.

specialist-council — AI Communication Intelligence

The Pi substrate already dispatches agents (dispatcher-core), runs single-domain production teams (marketing-orchestrator, video-director), debates one corpus (ceo-board, evidence-board), and gates output (opus-adversary, brand-guardian). What it lacks: any specialist actively pulling structured input from a DIFFERENT-domain specialist and integrating it before emitting. Hand-off today is one-way, upstream→downstream, via file artifacts. This skill adds the missing round-trip — the intelligence layer for agent-to-agent communication.

Core principle: specialists advise; the artifact's owner decides. The council gathers cross-domain input on a shared draft, folds it back, and bounds the loop. It never produces the artifact itself.

When to use

• A task whose quality depends on more than one specialist domain (content that must satisfy SEO + GEO + E-E-A-T + brand; a landing page; a video; a campaign).
• Inside an orchestrator's plan as a "wave 0: consult" step, or inside /spm as the specialist-board's cross-domain pass.

When NOT to use

• Single-domain work (a pure SEO audit → just call seo). No council overhead.
• As a build step — it critiques and enriches; it does not generate the deliverable.
• Replacing the single-domain orchestrators — it composes WITH them.

The protocol

1. Frame

Write the task + the current draft artifact to a shared file (council/{job_id}/draft.md + task.md). This shared corpus is what every specialist reasons over — reuse the evidence-board "N lenses over one corpus" primitive.

2. Route (which specialists to consult)

Pick from the routing table below by task type. Only pull specialists whose input would actually change the output — irrelevant consults are noise.

Task type	Consult (parallel)
Content / copy / blog / landing page	seo, geo-optimization, eeat (E-E-A-T), brand-guardian
Campaign / GTM / launch	marketing-orchestrator (owner), seo, geo-optimization, eeat
Video / explainer / reel	video-director (owner), heygen-director, remotion-designer, brand-guardian
Static / social visual — incl. "Canva"-class graphics, carousels, thumbnails, one-pagers	remotion-designer + design-board / design-canvas-html (layout), margot image_generate (imagery), brand-guardian — our-stack substitute for Canva; no new tool
Any client-facing FACT claim	source-ingest, eeat, brand-guardian

(Specialists marked "owner" produce; the rest consult. Add opus-adversary as a final adversarial gate for high-stakes output.)

3. Fan out (the communication step)

Dispatch the consult specialists in parallel via parallel-delegate (independent, over the same shared draft). Each is asked the SAME question: "Review this draft through your domain lens. Return structured input." Each returns a consult response (schema below) — never prose the owner has to parse.

Coordination mode (name it per task). Decision-grade / client-facing artifacts use centralized synthesis — the owner or /spm gathers all consult responses and integrates (high-bandwidth, one decision-maker). Reserve loose, independent fan-out (no synthesis round) for cheap, genuinely separable subtasks. Per DeepMind "From AGI to ASI" (arXiv:2606.12683, Pathway 4), centralized high-bandwidth coordination is what lets a specialist collective exceed any single agent; don't default to loose fan-out for work that needs a coherent result.

4. Synthesize + fold back

Merge the consult responses. Resolve conflicts explicitly (note disagreements, don't silently pick). Apply every must_fix; fold in suggestions where they don't conflict with the owner's intent. Record what was applied vs deferred.

5. Bounded re-consult

If applying the fixes materially changed the artifact AND any must_fix remains open, run ONE more round over the same specialists (mirrors opus-adversary's single loop-back). Never loop unbounded — kill-switch aware (~/.claude/HARD_STOP).

6. Emit

Return the enriched artifact + a consultation-record.json (who was consulted, their must_fix/suggestions, what was applied). The owner ships the artifact.

Consult-response contract (structured hand-off)

Every consulted specialist returns exactly this — the shared language of the communication layer:

{
  "specialist": "geo-optimization",
  "verdict": "pass | needs-work | fail",
  "must_fix": [{"issue": "no llms.txt / citable answer block", "why": "invisible to AI Overviews", "evidence": "geo-optimization §Trust"}],
  "suggestions": [{"change": "add an FAQ answer block for the head query", "impact": "high"}],
  "confidence": 0.0
}

Guardrails

• Bounded: max 2 consult rounds; parallel fan-out capped (parallel-delegate limits); kill-switch honored.
• Relevance over volume: consult only specialists whose input changes the output.
• Diversity over redundancy: collective gain comes from specialization diversity, not N copies of the same reviewer — roster genuinely distinct domain lenses and include at least one deliberately disconfirming one (DeepMind arXiv:2606.12683: collective intelligence depends on "diversity due to specialization"; mirrors the evidence-board opposing-source rule).
• Cooperation gate (anti-"solipsism"): each consulted specialist must cite and address at least one PEER's objection before its input is accepted — isolated, non-cooperative optimization is rejected. The paper names "solipsistic superintelligence" (agents optimizing in isolation) as the core failure mode; cooperation is a gate here, not a nicety. This is the council's oversight layer.
• Advise ≠ decide: the owner integrates; the council never overrides intent silently — conflicts are surfaced.
• Autonomy ladder: bounded single-domain consults run un-gated; cross-domain "group agent" strategic output (the ASI-shaped work) routes through human/Board review before shipping — maps to the existing policy-gate concern that multi-move execution is otherwise unguarded (see pi-dev-ops-autonomy-gate-layer).
• No new tools: consult only existing skills. The design/"Canva" lens is RESOLVED by substitution — Canva-class requests map to remotion-designer + design-board / design-canvas-html (layout) + margot image_generate (imagery), gated by brand-guardian (per feedback-skool-substitute-our-stack: install nothing new). E-E-A-T exists as eeat. Nothing stubbed here.

Integration

• Orchestrators: add a wave 0: consult entry to the wave-plan/production-brief before production waves, invoking this skill on the brief.
• /spm: the MOA bench (spm/references/moa-board.md, board_version 1.0+) embeds this skill's consult contract and cooperation gate in-memory rather than invoking the skill — spm's read-only constraint forbids the council/{job_id}/ writes. The contract is adopted verbatim there (plus slug/escalate fields); this skill remains the SSOT for the contract's semantics.
• dispatcher-core: expose as a skill.specialist-council step type.

Anti-duplication

Builds ON evidence-board (shared corpus), parallel-delegate (fan-out), dispatcher-core (dispatch), opus-adversary (loop-back). Does NOT replace the single-domain orchestrators, the board (decision-out), or the gates. It is the connective tissue between them — the enhance, not the rebuild.