Expert Panel — Painel de Especialistas Multi-Agente
diazhh/claude-expert-panel
Convoque um painel de especialistas reais que debatem até convergência
Ver no GitHubSkill para Claude Code que convoca um painel de especialistas (subagentes reais) para debater um tópico com acesso a ferramentas até chegarem a consenso. Cria automaticamente perfis de profissionais faltantes pesquisando a web. Ideal para decisões complexas, pesquisas e planejamento em equipe sintética.
Instalação
npx convez-ai diazhh/claude-expert-panel1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Avaliar decisões arquiteturais de software com múltiplas perspectivas
- Pesquisar abordagens técnicas comparando visões de profissionais distintos
- Revisar segurança, escalabilidade ou fluxos de autenticação com especialistas
- Gerar planos de projeto com input de diferentes disciplinas
Veredito

Implementação sofisticada de multi-agenting que substitui role-play por debate genuíno entre subagentes. Excelente para contextos que demandam síntese de múltiplas perspectivas técnicas, mas limitada ao ecossistema Claude Code e requer configuração manual. Ferramenta madura e bem documentada.
README
A tradução para o português deste README ainda está na fila.
Expert Panel — a Claude Code skill (/panel)
Convene a panel of specialist experts that debate a topic with real tool access until they converge on a deliverable — and fabricate any expert you don't have by researching the profession on the web.
A think-tank on demand — summon the profession, not the person.
Built for Claude Code. Each expert is a real subagent with its own context and read-only access to your codebase and the web. The main session acts as a facilitator that moderates the rounds and checks in with you at the key moments.
Why it's different
- Real multi-agent debate. Each expert is a separate subagent — not one model role-playing
many voices. They exchange positions through inter-agent messaging, cite evidence (
file:lineor URL), and genuinely disagree and change their minds. - Experts are fabricated on demand. Ask for a "corrosion engineer" you don't have and the skill web-researches the profession (standards, vocabulary, decision criteria, failure modes), writes a reusable Markdown profile, and seats them at the table.
- Two-tier, reusable experts. Generic roles live globally and work across every project; domain-specific roles live in your repo and are versioned with the code.
- Facilitator + human checkpoints. You approve the roster/agenda, break any deadlock, and sign off the final deliverable before it's written.
- Loop-until-dry convergence. The debate stops when a full round produces no new conflict and no new evidence — not after a fixed, arbitrary number of turns.
- Optional voice narration. On macOS, hear the debate read aloud with a distinct voice per
expert (
--narrar). Off by default. - Flexible output. The panel classifies your request and produces a research report, a decision record (ADR), or a plan.
Install
The folder must be named panel (it has to match the skill's name):
git clone https://github.com/diazhh/claude-expert-panel.git ~/.claude/skills/panel
Then restart Claude Code so /panel shows up in the slash-command menu. (You can also just
say "convene a panel about X" in natural language at any time.)
Requirements:
- Claude Code (the skill uses the
Agent,SendMessage, andAskUserQuestiontools). - Web access for fabricating new experts (
WebSearch/WebFetch). - macOS only for the optional voice narration (uses the built-in
saycommand).
Usage
/panel <topic> [--expertos N] [--rondas N] [--modo investigar|decidir|planificar] [--narrar] [--voz es_ES|es_MX]
Examples:
/panel is our event-ingestion architecture sound and scalable?
/panel evaluate our auth flow for security gaps --modo decidir --expertos 4
/panel research approaches for offline-first sync --modo investigar
/panel design the onboarding flow --narrar # read the debate aloud
| Flag | Default | Meaning |
|---|---|---|
--expertos N | auto (3–5) | Number of experts on the panel |
--rondas N | 3 | Substantive debate rounds (+1 deadlock round) |
--modo | auto | Force the deliverable: investigar / decidir / planificar |
--narrar | off | Read the debate aloud (macOS say, one voice per expert) |
--voz | es_MX | Narration accent: es_ES or es_MX |
User-facing prompts are written in Spanish (es-VE) by default — adapt the reference files if you want another language.
How a session runs
- Intake + roster (checkpoint). The facilitator classifies the deliverable mode, proposes a roster of experts plus an agenda, and asks you to approve or adjust.
- Resolve the roster (two-tier). Existing profiles are loaded (project → global). Any missing expert is fabricated by web research, with a scope checkpoint before saving.
- Convene + debate. Each expert is spawned as a named subagent with read-only tools. Round 0 collects opening positions; later rounds relay conflicts between experts until loop-until-dry or a deadlock (which you break).
- Synthesize + persist. The deliverable is written in the chosen mode; the transcript and
deliverable are saved to
<repo>/.claude/panel-sessions/<date>-<slug>/.
Experts
An expert is a Markdown file with YAML frontmatter (its dossier) plus a body that becomes the
subagent's system prompt. The schema is in templates/expert-profile.md:
---
name: process-engineer
role: Industrial Process Engineer
scope: global # global | project
specialty: …
researched: 2026-06-13
sources: [https://…]
expertise:
standards: [ISA-95, IEC 62443]
watches_for: [unclosed control loops, calibration drift]
decision_criteria: [process safety, traceability]
persona: conservative on safety; cites standards; distrusts shortcuts
---
# body = the expert's system prompt (identity, how they reason, what they watch, hard rules)
Resolution is two-tier: the panel looks in <repo>/.claude/experts/<name>.md first, then in
~/.claude/skills/panel/experts/<name>.md. If neither exists, it creates one (see
references/create-expert.md).
Included generic experts
These ship with the skill (global scope) and are reusable on any project:
| Expert | Reasons with |
|---|---|
software-architect | C4, ADR, DDD, SOLID, module boundaries, reversibility |
process-engineer | ISA-95/88, IEC 62443, SCADA, material balances |
senior-developer | TypeScript, clean code, TDD, technical debt, DX |
ui-ux-designer | Nielsen heuristics, WCAG 2.2, ISA-101, design systems |
product-manager | RICE / MoSCoW / Kano, OKRs, North Star, discovery |
business-analyst | BABOK v3, BPMN 2.0, INVEST, requirements traceability |
qa-test-engineer | ISTQB, the test pyramid, TDD/BDD, mutation testing |
devops-sre | Google SRE, DORA, IaC/GitOps, observability, SLOs |
technical-writer | Diátaxis, docs-as-code, minimalist technique |
Domain-specific experts (e.g. a custody-transfer-standards expert for an oil-and-gas project) are created on demand and saved into your repo, not shipped here.
Output modes
See references/output-modes.md:
- investigar → a cited, multi-perspective research report.
- decidir → an ADR: recommendation + options + risks + assumptions + dissent.
- planificar → a CONTEXT brief ready to feed a planning workflow.
Every deliverable ends with the panel's confidence (high / medium / low) and why.
Voice narration (optional, macOS)
With --narrar, the facilitator reads the debate aloud through the macOS say command, assigning
a distinct Spanish voice to each expert. It narrates a short spoken summary per turn and reads the
final recommendation in full. Audio is a pure side-channel — the written transcript is always the
source of truth. Details in references/narration.md.
Cost
Real subagents are powerful but not free. The facilitator enforces conservative defaults: 3–5
experts, ~3 rounds, experts capped at ~200 words per turn, and web research only when creating an
expert (after that, the profile is reused). Tune with --expertos / --rondas.
Repository layout
panel/
SKILL.md # entry point (what Claude Code loads)
README.md # this file
DESIGN.md # the design rationale
references/
run-meeting.md # the end-to-end session protocol
facilitator.md # convergence, deadlock, checkpoint rules
create-expert.md # fabricating a missing expert via research
output-modes.md # the three deliverable templates
narration.md # optional voice narration
templates/
expert-profile.md # the expert profile schema
experts/ # the included generic experts
Credits
Designed and built with Claude Code using the superpowers brainstorming → writing-plans → subagent-driven-development workflow. The multi-agent debate pattern is inspired by BMAD's "party mode".
License
MIT.