Claude Council Skill · Research and review

Claude Council Skill: Compare Independent Claude Perspectives

Four perspectives, one better-informed decision.

Amgad El Gamal’s Claude Council Skill runs independent Claude subagents in fresh contexts, anonymously peer-reviews their answers, and lets the main session synthesize consensus and dissent. It runs on the Claude Code subscription you already have; this page is for amgadelgamal/claude-council, separate from Karpathy’s llm-council application.

View Amgad El Gamal’s source

Claude perspectives
Anonymous peer review
Council synthesis

About This Skill

Claude Council Skill is Amgad El Gamal’s installable workflow for asking independent Claude subagents for perspectives on one question, anonymously peer-reviewing the answers, and returning a synthesis with consensus and dissent. It runs on Claude Code and is distinct from Karpathy’s llm-council app.

Built on proven foundations

Trusted Source, Professionally Refined

The original expertise stays intact. Focused refinements make it easier to start, follow the guidance, and know when the work is finished.

Original foundation

Amgad El Gamal · commit 819c1d0

The original Skill knowledge, package structure, attribution and original license terms are preserved. Follow the source link to see where this edition began.

View the original project

Antigravity optimization

4 files, one clearer workflow

A source-faithful edition with the improvements explained below. The download includes the original attribution and everything needed to follow the Skill’s guidance.

More useful from the first prompt

What We Improved

Keep the depth you came for. Get clearer starting points, practical guidance and an explicit finish line.

A focused council trigger and entity boundary

Choose a second-opinion workflow built from independent Claude perspectives, with Amgad El Gamal’s Skill kept separate from Karpathy’s application.

A brief that protects the question boundary

Give independent members one shared question, decision boundary, and context before they form separate opinions.

An explicit synthesis finish line

Leave the review with a named plan, consensus, dissent, and attribution as the handoff fields.

Start with One Question

1State one decision or question, the context each council member needs, and the output you want from the review.
2Invoke `/claude-council` or ask to consult the council; the default seats use independent Claude subagents in fresh contexts.
3In full mode, let the members answer in parallel and anonymously peer-review one another; quick mode skips peer review.
4Use the main session’s synthesis to separate consensus, dissent, attribution, and unresolved uncertainty.

Example Command

Consult the council: review this architecture decision and return the strongest plan, consensus, dissent, and attribution.

Core Capabilities

Fresh Claude perspectives

Put one question in front of independent Claude subagents, each working from a distinct persona and fresh context.

Anonymous peer review

Have the members critique and rank anonymized responses without anchoring on another member’s name.

Synthesis with consensus and dissent

Let the main session weigh the reviews and present a plan with attribution, consensus, and dissent.

Claude Code subscription

Run the council through the Claude Code subscription already in use rather than adding a separate vendor API key.

Use Cases

Architecture review

Compare independent perspectives before choosing a technical direction.

Research question

Ask independent Claude perspectives to examine a question, then inspect what they agree or disagree about.

Prompt or plan critique

Use a council to surface blind spots in a proposed prompt, plan, or decision.

SKILL.md

# Claude Council

This page is bound to Amgad El Gamal's MIT-licensed Claude Council Skill at commit 819c1d07aace94b4bd0c43ccc316e864ec62962a. It is a root SKILL.md plus its documented scripts/workflow, not Karpathy's separate llm-council application or a generic multi-agent bundle.

The source job is a three-stage review: independent Claude subagents answer in parallel through distinct personas and fresh contexts; full mode anonymously peer-reviews and ranks those answers; the main session synthesizes a plan with attribution, consensus, and dissent. Members run on the user's Claude Code subscription, with no external API keys. This page records the source workflow, not a live council result.

Grill MeA separate workflow for pressure-testing plans and decisions with the user.SuperpowersA separate collection for structured implementation work.

Frequently Asked Questions

What is Claude Council Skill, and how do I install it?

It is Amgad El Gamal’s installable Claude Code Skill. Clone `https://github.com/amgadelgamal/claude-council` into `~/.claude/skills/claude-council` (or your project’s `.claude/skills/claude-council`), start a new session if needed, then invoke `/claude-council` or ask to consult the council.

Is this Karpathy’s llm-council app?

No. Karpathy’s llm-council is a separate method/application reference. This page is bound to amgadelgamal/claude-council: its members are Claude subagents on the user’s Claude Code subscription, not a cross-vendor council that requires several API keys.

How many models does the council use?

The default council has four persona seats—Architect, Skeptic, Pragmatist, and Researcher. A custom council can use 2–8 members; quick mode skips the anonymous peer-review stage.

Does Claude Council guarantee consensus?

No. Its useful boundary is comparison and synthesis. Disagreement, not included providers, and unresolved uncertainty should remain visible.

What should I give Claude Council?

Give it one decision or question, the relevant project paths or context, the output format you want, and the constraints that should shape the review. Members can inspect the working directory when the question is about a project.

Discussion

0 comments

Trigger Phrases

Use these phrases to activate this skill in your AI coding assistant:

Use Claude Council for this decision.Ask multiple agents to review this question.Compare several model perspectives before we choose.