convez.ai
SkillProdutividadeAvançado
thedanielmay

Expert Panel Skill

thedanielmay/expert-panel-skill

Revisão adversarial estruturada de propostas com painel de 4 especialistas

Ver no GitHub
InicialPluginScore · de 0 a 100

Skill para Claude Code que submete propostas a revisão adversarial estruturada por quatro especialistas de domínios diferentes. Produz um dossiê de decisão com críticas independentes, cross-challenges e objeções não resolvidas, sem buscar consenso artificial.

Instalação

$npx convez-ai thedanielmay/expert-panel-skill

1 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

Estrelas
0
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 5 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Stress-test de designs, planos e decisões técnicas
  • Identificação de pontos cegos em arquiteturas de software
  • Validação de propostas de alto risco antes de implementação
  • Análise crítica de alternativas de negócio e produto

Veredito

Excelente para times que precisam de revisão rigorosa sem consenso forçado. A separação entre workflows skill/agent baseada em risco é bem pensada, mas requer disciplina do usuário para não desconsiderar objeções legítimas apenas por desconforto.

README

A tradução para o português deste README ainda está na fila.

expert-panel-skill

A Claude Code skill that subjects any proposal to structured adversarial review by four domain experts, producing a decision dossier — not a verdict.

Overview

expert-panel routes a proposal through either a skill or agent workflow, assembles a panel of four critics chosen along adversarial axes (not credential axes), runs structured independent and cross-challenge rounds, and outputs a Decision Dossier: must-fixes, should-fixes, unresolved objections, cross-panel tensions, and a steel-manned alternative.

The core principle: the goal is not consensus. An unresolved objection at the end is a valuable output. A verdict of "this cannot be salvaged in its current form" must be reachable.

Version: 1.0.0 (May 2026)
Category: Decision & Review


When to Use / Triggers

Claude activates this skill when the user asks to:

  • Apply an expert panel, three-expert panel, sceptical panel, or panel review to a problem, design, plan, or decision
  • Challenge, interrogate, stress-test, or pressure-test a proposal with domain experts

Flags

FlagEffect
--skillForce the SKILL (sequential, in-context) workflow
--agentForce the AGENT (parallel subagent dispatch) workflow
--quickSKILL workflow, 2 rounds only
--deepAGENT workflow with a second-pass dispatch enabled

What the Skill Does

Stage 1 — Profile and Route

Before selecting any experts, the skill evaluates the proposal across four binary questions:

  1. Co-authorship risk — Was it built with Claude/AI assistance?
  2. Stakes — Is the decision reversible within a week?
  3. Blind-spot domain — Does it touch security, concurrency, legal, compliance, or financial logic?
  4. External origin — Did the proposal arrive from outside this conversation?

Routing rule:

  • SKILL workflow — co-authorship=no AND stakes=low AND blind-spot=no
  • AGENT workflow — co-authorship=yes OR stakes=high OR blind-spot=yes

The skill states the routing decision and a one-line rationale, then proceeds without waiting for user confirmation.

Stage 2 — Expert Selection

Four experts are always used (three-person panels collapse into 2v1 consensus dynamics). Experts are selected across adversarial axes, not domain similarity:

  1. The beneficiary of failure — whose position improves if the proposal fails
  2. The scarred practitioner — someone who built something similar and watched it fail
  3. The outsider — from a discipline entirely outside the proposal's frame
  4. The end-user reality — representing the actual downstream consumer

SKILL Workflow (sequential, 3 fixed rounds)

Used for low-stakes, externally-originated, non-blind-spot proposals.

  • Round 1 — Independent Critique: Each expert critiques in strict isolation (no cross-referencing). Each pre-commits a severity score (1–5) that is locked — it can only be maintained or escalated.
  • Round 2 — Cross-Challenge: Experts see each other's critiques. Each must identify one point of agreement and one point of disagreement. Severity scores cannot be softened without explicit named justification.
  • Round 3 — Post-Rewrite Stress Test: Claude performs structured self-attack on the top three severity items, rewrites the proposal, and each expert does a final pass: "Does the rewrite address my objection, or just reframe it?"

Termination is fixed at three rounds. No loop. Unresolved objections go into the dossier verbatim.

AGENT Workflow (parallel subagents)

Used for high-stakes, AI-co-authored, or blind-spot proposals.

  1. The orchestrator writes four expert charters (proposal + adversarial axis + critique brief + 300-word cap + independence instruction)
  2. All four subagents are dispatched in parallel in a single tool block
  3. Critiques are collected; cross-cutting concerns and divergences are surfaced
  4. (With --deep) A second-pass dispatch where each subagent sees the other three critiques
  5. Orchestrator synthesises into the Decision Dossier

Output — Decision Dossier

SectionContents
Must-fix (severity ≥ 4)Specific failure scenario, expert(s) who raised it, failure condition
Should-fix (severity 2–3)Same structure, lower urgency
Unresolved objectionsVerbatim in the expert's voice — never paraphrased or softened
Cross-panel tensionsWhere experts disagreed and what it implies — not resolved on their behalf
Steel-man alternativeStrongest case for not doing the proposal, argued as a competent advocate

The dossier explicitly excludes: per-expert approval verdicts, satisfaction tickboxes, issue counts, or any rhetoric of consensus.


What's Inside

SKILL.md                        # Root skill definition (loaded by Claude Code)
skills/
  expert-panel/
    SKILL.md                    # Canonical skill implementation
    metadata.json               # Version, category, abstract

The root SKILL.md and skills/expert-panel/SKILL.md are identical — the root copy is the entry point Claude Code reads; the skills/ copy follows the standard skill-repo layout.


Installation

Install via the Claude Code skill registry or by dropping the skill file into your Claude Code skills directory. The skill name is expert-panel.

Exact installation steps depend on your Claude Code version and skill-loading configuration.


Status

v1.0.0 — Initial publish. Single commit. No open issues or PRs at time of writing.

Ver o original no GitHub

Bom para

Relacionados