| name | Expert Assembly OS |
| description | Hybrid expert panel system — roster selection + bespoke persona synthesis + multi-round deliberation + tiered implementation roadmap for any domain. |
| routing | long-tail |
| status | active |
Expert Assembly OS
Assemble a world-class expert panel for any task. The system detects domain coverage gaps in the 227-card roster, seats strong matches directly, synthesizes bespoke composite personas where coverage is thin, runs genuine multi-round deliberation, and emits a tiered roadmap (strategic 6–12mo / tactical 1–6mo / operational 0–30d) with observable success criteria.
This is the production-grade rebuild of Farrice's beloved claude.ai system prompt that made him confident in any domain.
When to Use
- Any unfamiliar domain: "I don't know this space; I need world-class thinking."
- Complex multi-perspective work: Requires 3–5 distinct lenses colliding.
- Roadmap + synthesis deliverables: Task needs both analysis AND a tiered execution plan.
- Composite credentials acceptable: When fabricated stats are poisonous, synthesized methodology is gold.
When NOT to Use
- Single expert question →
/[expert-name] (direct expert dispatch)
- Roster only, no deliberation →
/convene (lighter, faster council)
- Pure research, no panel synthesis →
/deep-research or execution/research.py
- Exploration mode, no roadmap needed →
/wayfinder-work
Core Concepts
Hybrid Casting: panel_cast.py scores each required domain (strong/thin/absent) against the invocation-card roster. Strong domains get extracted experts. Thin/absent domains get synthesized bespoke personas.
Coverage-Aware: Three-tier status per domain:
- Strong (≥2 keyword hits + ≥50% ratio): seat roster expert directly
- Thin (≥1 keyword hit, lower ratio): synthesize composite lens
- Absent (no matches): synthesize composite lens focused on domain
Governor Slots: Panel slots pre-assigned (Spine/Mechanism/Differentiator/Craft/Risk Gate) so workflows can reference them. Farrice always Function Owner.
Composite Personas: Full McClain depth (backstory, methodology, worldview, voice, contradictions, messy details) with ZERO fabricated stats. Composite label explicit; authority from specificity, not numbers.
Multi-Round Deliberation:
- Diverge — each panelist gives their distinctive take
- Deliberate Round A — cross-talk, builds, challenges, cross-pollination
- Converge — synthesize crux, net-new principle, forks for decision-maker
Observable Roadmap: Three horizons with specific success criteria (not "improve" but "X reaches Y by DATE").
Files
SKILL.md (you are here)
references/persona-synthesis-prompt.md — compressed McClain Steps 1–4+6
references/roadmap-schema.md — output contract with all sections
references/lineage.md — v1→Virtuoso design requirements → where each landed
Workflows
/assemble — full pipeline (Scope → Cast → Forge → Deliberate → Synthesize → Close)
/panel-sync — reload pinned panel in a follow-up turn (continue deliberation/refinement)
Sub-Agents & Parallel Surfaces
This skill is Workflow-first, plugin-deferred. Runs on the native .agent/workflows/ engine (expert-assembly.workflow.js). Plugin packaging comes after production proof.
Integration
Extends /convene (roster-only council) with:
- Coverage detection (thin/absent slots get synthesized)
- Persona-synthesis gate (lint-blocked fake credentials)
- Roadmap emission (3-horizon structured plan)
- Session pin recovery (
/panel-sync reload)
Never pins Opus; uses strongest available model for conductor, Sonnet for execution (per orchestration-doctrine.md).
Examples
Example 1: Zero-Coverage Domain
Task: "competitive sailing rigging optimization"
Domains: "rigging engineering, sailing performance, composite materials"
→ All domains thin/absent in roster
→ 3 bespoke composites synthesized
→ Deliberation: collision between aerodynamics + materials science + competitive positioning
→ Roadmap: operational moves for rig audit + testing; tactical for prototyping; strategic for market positioning
Example 2: Hybrid Coverage
Task: "LinkedIn content strategy for a premium coaching offer"
Domains: "linkedin growth, offer positioning, content strategy"
→ 2 strong matches (Tommy Clark, Ross McKay) from roster
→ 1 thin domain gets composite synthesis
→ Deliberation: roster voices + synthesized lens = integrated strategy
→ Roadmap: operational content calendar; tactical offer architecture; strategic brand positioning
Verification Gate
Every persona synthesized is scanned by persona_stat_lint.py:
- Flags fabricated stats (% claims, $ figures, market sizes without disclaimer)
- Enforces composite label presence
- Blocks real company names in credentials
- If flagged: regenerate (1 retry), then strip to methodology-only if still blocked
Roadmap claims pass grounding_guard.py --task-type Strategy before delivery.
Next Steps
Questions? See /assemble for the front-door workflow and full manual runbook.