Council of the Wise

A Clawdbot skill that spawns sub-agents to analyze ideas from multiple expert perspectives.

What It Does

When you say "send this to the council" or "council of the wise," Clawdbot:

1. Sends a loading message: 🏛️ *The Council convenes...* (this takes 2-5 minutes)
2. Spawns a sub-agent that analyzes your idea through multiple lenses:

• 👹 Devil's Advocate — Challenges assumptions, finds weaknesses
• 🏗️ Architect — High-level structure and design
• 🛠️ Engineer — Implementation feasibility and details
• 🎨 Artist — Voice, style, and user experience

The sub-agent returns a synthesized report with insights, concerns, and recommendations from each perspective.

Installation

Via ClawHub

clawhub install council-of-the-wise

Manual

Copy the council/ folder to your Clawdbot skills directory:

cp -r council/ ~/clawd/skills/council/

Usage

"Send this to the council: [your idea]"
"Council of the wise: [topic to analyze]"
"Get the council's feedback on [thing]"

Example

Input:

Send this to the council: I'm building a CLI tool that uses AI to help with bug bounty hunting — automated recon, AI-assisted analysis, human-in-the-loop for key decisions. Target: HackerOne programs.

Output:

## 🏛️ Council of the Wise — AI Bug Bounty Tool

### 👹 Devil's Advocate
I've seen three ways this dies:

1. **Token costs eat your bounties.** Running GPT-4 on every subdomain? 
   You'll spend $50 in tokens to find a $100 bug. Do the math first.
2. **HackerOne ToS.** "Automated scanning" is a gray area. One aggressive 
   program bans your account, you lose access to all of them.
3. **The 80% problem.** AI is great at recon. It's terrible at the creative 
   leap that finds the actual vulnerability. You'll automate the easy part 
   and still do all the hard work manually.

Path forward: Start with ONE program. Track tokens religiously. Keep human 
approval before any active testing.

### 🏗️ Architect  
The fundamental constraint is the human-AI handoff. Structure it as three layers:

1. **Recon Layer** (fully automated): subdomains, ports, tech stack
2. **Analysis Layer** (AI-assisted): "here's what looks interesting and why"
3. **Testing Layer** (human-approved): actual vulnerability validation

The AI should be a force multiplier for your attention, not a replacement. 
Design the handoff points explicitly.

### 🛠️ Engineer
Concrete build plan:

- **Week 1:** Scope parser + subfinder + httpx pipeline (no AI yet)
- **Week 2:** Add Claude for analyzing httpx output, flag interesting endpoints
- **Week 3:** Human-in-the-loop approval flow, Slack/Telegram notifications
- **Week 4:** Cost tracking dashboard, per-program budgets

Start with tools you know: Python, existing recon tools, Claude API. 
Don't build a framework — build a workflow.

### 🎨 Artist
The experience should feel like a skilled research assistant, not a robot:

- Notifications should sound human: "Found 3 interesting endpoints on 
  target.com — the /api/v2/admin looks promising, want me to dig deeper?"
- Let the user name campaigns, add notes, build a narrative
- The tool should remember context: "Last time on this program, you found..."

### ⚖️ Synthesis
**Do it, but small.** The council agrees on:
- Start with ONE HackerOne program
- Track token costs from day one
- Keep human approval for anything active
- Design for attention-multiplication, not automation

**Key decision:** Build for yourself first, or build for others? 
Recommendation: yourself first, for 3 months minimum.

**Where the council disagreed:** Engineer wants to ship fast with existing tools. 
Architect wants to design the human-AI handoff carefully first. Both valid — 
depends on whether you're exploring or committing.

---
📊 **Token Usage:** ~2.1k input / ~1.8k output tokens *(estimated)*

Adding Custom Agents

Just drop a .md file in the agents/ folder:

# Add a security-focused reviewer
cat > agents/Pentester.md << 'EOF'
# Pentester

*"What's the attack surface? How would I break this?"*

You think like an attacker. Every feature is a potential vulnerability...
EOF

The skill auto-discovers all agents — no config changes needed.

Agent File Schema

Agent files should follow this minimal structure:

# Agent Name

*"A memorable quote or philosophy"*

[1-2 sentences describing the perspective]

## Your Voice
[How this agent sounds — tone, example phrases]

## Your Approach
[How this agent analyzes — what they look for]

See the bundled agents in agents/ for examples.

Custom Agents

The skill includes bundled agent personas, but if you have custom PAI agents at ~/.claude/Agents/, those will be used instead.

HOWTO: Using Council of the Wise Effectively

When to Use the Council

• Before committing to a major decision (new project, pivot, launch)
• When you're too close to an idea and need outside perspective
• For stress-testing plans before sharing with stakeholders
• When you're stuck and want structured thinking prompts

When NOT to Use It

• Quick questions ("what's the syntax for X?")
• Time-sensitive tasks (takes 2-5 minutes)
• Small decisions where four perspectives is overkill
• Things you've already decided — the council isn't for validation

Getting the Best Results

1. Be Specific. "Analyze my startup idea" → weak. "Analyze this B2B SaaS for security teams: [specific pitch]" → strong.

2. Include Context. Share constraints, goals, and what you've already considered. The council is smarter when you're honest about what you don't know.

3. Ask Follow-ups. After the council reports, dig into the most interesting points. "The Devil's Advocate mentioned X — expand on that."

4. Use the Synthesis. The individual perspectives are interesting; the synthesis is actionable. Start there if you're short on time.

Example Invocations

"Send this to the council: I'm considering switching from Substack to 
Beehiiv for my newsletter. 2000 subscribers, mostly free, want to 
monetize. What should I consider?"

"Council of the wise: Review this README before I publish it — 
is the value prop clear? What's missing?"

"Get the council's feedback on this feature spec [paste spec]"

Pro Tip: Run the council before you're emotionally invested in an idea. It's easier to hear criticism early than after you've spent a week building.

Why These Four Perspectives?

The council members complement each other:

• Devil's Advocate finds what could go wrong (risk)
• Architect designs how it should be structured (strategy)
• Engineer figures out how to build it (execution)
• Artist shapes how it should feel (experience)

Together they cover: risk, strategy, execution, and experience — the four dimensions most ideas need to succeed

Credits

Inspired by Daniel Miessler's PAI (Personal AI Infrastructure) concepts. The Architect, Engineer, and Artist agent personas are adapted from PAI patterns. The Devil's Advocate is an original creation.

License

MIT