/ 目录 / 演练场 / Carmack-Council
← 全部服务器 ↗ GitHub
● 社区 SamJHudson01 ⚡ 即开即用

Carmack-Council

作者 SamJHudson01 · SamJHudson01/Carmack-Council

Multi-agent Claude Code framework: 10 expert sub-agents (Carmack, Fowler, Kent Beck, Troy Hunt...) each review code in parallel with their own 200k context.

The Carmack Council structures development as spec → plan → implementation → review, with feedback loops. The headline innovation is parallel expert subagents: the Chair orchestrator dispatches 10 specialists simultaneously (Carmack engineering, Fowler refactoring, Kent Beck testing, Troy Hunt security, Matteo Collina backend, Brandur Postgres, Vercel perf, Simon Willison LLMs, Karri Saarinen UI, Vitaly Friedman UX). Each gets its own 200k context for its domain.

▶ 观看演示 📦 安装 💡 使用场景

为什么要用

核心特性

实时演示

实际使用效果

carmack-council-skill.replay ▶ 就绪
0/0

安装

作者仓库 · Claude Skills 官方文档

使用场景

实战用法： Carmack-Council

Get a 10-expert review on a PR before you merge it

👤 Engineers on small teams without a deep review bench ⏱ ~30 min intermediate

何时使用： You'd like a security person, a perf person, a testing person to look at your PR — but there's just you.

步骤
  1. Invoke the review stage
    Carmack Council review: PR <link or diff>. Dispatch all experts.✓ 已复制
    → Ten parallel reviews running, each domain-scoped
  2. Read the merged report
    Summarize the top 3 issues across all experts, ordered by severity.✓ 已复制
    → Prioritized, not 10 disjoint reports
  3. Fix and re-review
    I fixed 1 and 2. Re-review with just Troy Hunt and Brandur for issue 3.✓ 已复制
    → Focused re-review, not full re-dispatch

结果： A PR that's survived 10 expert lenses — stronger than a rubber stamp from a teammate.

注意事项
  • Experts disagree — The Chair merges and prioritizes — if two experts conflict, pick the one your situation weights heavier
搭配使用： github

Run a feature through the full spec → plan → implement → review pipeline

👤 Small teams building a feature end-to-end with AI ⏱ ~120 min advanced

何时使用： Feature is substantial and you want structured gates.

步骤
  1. Spec
    Carmack Council: spec for adding Stripe-based subscriptions.✓ 已复制
    → Carmack-opinionated spec — pragmatic, not architecture-astronautic
  2. Plan + implement
    Proceed to plan, then implement in the smallest vertical slice.✓ 已复制
    → Thin working slice, not a scaffolding swamp
  3. Council review
    Review with all 10 experts.✓ 已复制
    → Cross-domain findings tied to actual lines

结果： A feature shipped with a thoroughly reviewed implementation.

注意事项
  • You treat expert style as gospel — They're opinions backed by real engineers' writing — they're evidence, not commandments
搭配使用： github · planning-with-files

组合

与其他 MCP 搭配，撬动十倍杠杆

carmack-council-skill + github

Run council_review against a real PR, post summary as a PR comment

Council review PR #47; post the prioritized top-5 as a comment.✓ 已复制
carmack-council-skill + planning-with-files

Track council findings across multiple review rounds

Use planning-with-files progress.md to track which council findings are resolved.✓ 已复制

工具

此 MCP 暴露的能力

工具输入参数何时调用成本
spec_stage feature intent Start of a feature 0
plan_stage approved spec After spec 0
implement_stage approved plan Actually building 0
council_review diff or PR Before merge High — 10 parallel agents
single_expert_review diff + expert name Re-check a specific concern after fix 1/10 of full council

成本与限制

运行它的成本

API 配额
None
每次调用 Token 数
High — council_review runs 10 parallel agents
费用
Free — skills are local files. LLM token cost scales with parallelism.
提示
Use single_expert_review for targeted re-checks; full council is expensive per run.

安全

权限、密钥、影响范围

凭据存储： No credentials — pure prompts
数据出站： None — all review happens in-session

故障排查

常见错误与修复

Wrong trigger keywords — only one expert activates

Say 'dispatch all experts' or 'full council' explicitly.

替代方案

Carmack-Council 对比其他方案

替代方案何时用它替代权衡
Single-reviewer vanilla ClaudeSmall changes where 10 experts is overkillCheaper, less coverage
Human code reviewTeam has real reviewersBetter but slower — Council complements, doesn't replace

更多

资源

📖 阅读 GitHub 上的官方 README

🐙 查看未解决的 issue

🔍 浏览全部 400+ MCP 服务器和 Skills