Challenge your thinking before you commit.

"Ready to decide" plan
↓
4 objections surfaced
2 unverified assumptions / 1 cheaper path / 1 hidden downside
↓
Decision revised
↓
Committed with fewer regrets

The riskiest decisions are the ones that feel obvious too early.
Before You Decide runs your plan past an independent advisory panel and tells you what they would challenge — before reality does.

Because the question is not:
"Does this plan make sense?"
The question is:
"Who would disagree — and why?"

Council — LLM 2-Team Decision Path Skill

Summary (EN)

Karpathy-style LLM Council restructured as a 2-team decision support skill. 13 advisor roles (6 Business + 4 Technical + 3 Floating) produce independent perspectives on any single question or proposal. Anonymous peer review surfaces blind dissent. Chairman synthesis includes user-goal reverse calculation. Pure local template generation — zero external API cost, no LLM dependencies.

サマリ (JP)

Karpathy 発案 LLM Council 概念を、エージェント向けに 2チーム制（ビジネス／技術）の意思決定支援スキルとして再構築。13名のアドバイザー役（ビジネス6名＋ 技術4名＋ フローティング3名）が独立視点で1つの問い／提案を多角評価。匿名ピアレビュー（任意）で盲点を炙り出し、Chairman 統合判断には user_goal 逆算チェックを含む。ローカルテンプレ生成のみ・外部API課金ゼロ・LLM依存なし。

Target Users

• Solo founders / indie consultants making strategic decisions alone
• Small teams without a dedicated decision-review process
• Content / product creators evaluating pricing, positioning, new initiatives
• AI agent builders wanting structured multi-perspective synthesis without per-call LLM cost
• Anyone who's noticed "one LLM gives one perspective" is not enough for important choices

Why Council

Architecture

Question / proposal
   │
   ▼
council.py ─── config/council_config.json (7 placeholders + role definitions)
   │
   ├── Team Selection (auto / A / B / mixed)
   │
   ├── Active advisors produce independent views:
   │     - A1 Contrarian / A6 Marketer / A7 CFO / A10 Entrepreneur
   │     - A11 Operations / A15 Tech Lead / A5 Executor / A12 Cautious
   │     - A13 SysEngineer / A14 ProductDev / A4 Outsider (always)
   │     - A2 First Principles / A3 Expansionist (on demand)
   │
   ├── [Optional] Anonymous peer review pass
   │
   ▼
Chairman Synthesis (markdown report)
   - Main disagreements
   - {user_goal} reverse calculation
   - Recommendation
   - Single next action
   - Caveats

Quick Start

# 1. Extract skill bundle
unzip council_v1.0_skill.zip
cd council_v1.0_skill

# 2. Try on bundled sample question
python3 scripts/council.py \
  --question "Should we launch a paid newsletter at $5/month, or stay free?" \
  --team auto \
  --output-dir ./out

# 3. Read the council report
cat out/council_report_*.md

Customization (7 placeholders)

Edit config/council_config.json:

{
  "active_preset": "solo_founder",
  "presets": {
    "solo_founder": {
      "user_name": "the founder",
      "user_role": "solo decision-maker",
      "user_profile": "one-person operation with limited time and bootstrapped budget",
      "user_context": "early-stage product / content building",
      "user_constraints": "5-10 hours per week of strategic time, $1000/month operating budget",
      "user_goal": "$10,000 MRR within 12 months",
      "audience_persona": "target audience the user is building for"
    }
  }
}

Bundled presets:

• generic (default — neutral context)
• solo_founder (one-person operation)
• corporate_team (mid-size team with quarterly OKR rhythm)
• creator_economy (content creator, audience-focused)

Team Selection

A4 (Outsider with {audience_persona} view) is always present. A2 (First Principles) and A3 (Expansionist) summon on demand.

Anonymous Peer Review (optional pass)

python3 scripts/council.py \
  --question "Quarterly direction: B2B or B2C?" \
  --team A \
  --peer-review \
  --output-dir ./out

After all advisors produce their views, the script shuffles them and asks each advisor: "Which view is most insightful, and why?" This surfaces dissent that majority-bias would otherwise hide.

What's Included

council_v1.0_skill/
├── LICENSE                    # MIT
├── README.md                  # quick start
├── SKILL.md                   # Capafy standard skill manifest
├── agent-card.md              # this file
├── config/
│   └── council_config.json    # 7 placeholders + role definitions
├── scripts/
│   └── council.py             # main entry (Python 3.10+, stdlib only)
└── sample/
    ├── input/sample_question.txt
    └── output/
        ├── expected_council_report.md
        └── expected_log.jsonl

Compatible Agent Clients

• Claude Code (CLI invocation + Python API)
• OpenClaw
• Codex
• Anthropic Claude Agent SDK
• Standalone CLI (no agent required)

Differentiation vs. "Ask one LLM"

The Council skill produces structured templates for each role's view. You can either:

1. Fill them yourself (manual reflection mode — zero LLM cost)
2. Pipe each role's template into your own LLM call (parallel-agent mode — your control over cost)

Either way, the structure (13 roles, 2 teams, peer review, goal reverse calculation) is the value — the skill ensures you don't skip the structured-thinking steps that single-LLM calls flatten.

Origin

Inspired by Andrej Karpathy's LLM Council concept (multi-LLM-as-advisors pattern), restructured for:

• Agent-friendly 2-team decomposition (Business vs Technical)
• Explicit {user_goal} reverse calculation in Chairman synthesis
• Anonymous peer-review option to surface blind dissent
• Local-only templating to remove per-decision API cost

License

MIT (see LICENSE). Free for commercial use. Modify and redistribute permitted.

Support

• Documentation included in SKILL.md and README.md
• No vendor lock-in — works offline after one-time download
• Customizable per content domain via 7 placeholders + role definitions