godmode

内容来源：README.md（说明文档） · 原始地址 · 查看安装指南

原始内容

# GODMODE — AI Development Team
### Powered by NEURON.ONE Architecture

> One person. Many projects. Corporate power in one pair of hands.

GODMODE turns Claude Code into a full AI development team. Multi-agent orchestration, cross-model disputes (Claude + GPT + Gemini), 50+ skills, smart model routing — all configured and ready to use.

Built by a non-programmer using AI. If I can do it, so can you.

---

## What You Get

- **16 Orchestrator Rules** — battle-tested operating principles for AI-assisted development
- **NEURON.ONE Architecture** — adaptive swarm: one orchestrator + dynamic agents created per task
- **Cross-Model Disputes** — Claude, GPT, and Gemini debate until consensus on architecture decisions
- **50+ Skills Library** — security, development, design, testing, infrastructure, and more
- **Smart Model Routing (Ecomode)** — Haiku for simple tasks, Sonnet for medium, Opus for complex
- **Agent Communication Protocol** — JSON envelopes, preflight/postflight validation, atomic writes
- **Memory System** — hot/warm/cold tiers with auto-persistence across conversations
- **External Advisors** — GPT for security/architecture, Gemini for design/visual analysis

## Architecture

<p align="center">
  <img src="docs/architecture.svg" alt="GODMODE Architecture" width="100%"/>
</p>

**Level 1 — Orchestrator**: permanent Claude instance that understands the full picture, knows all skills, and decides how many agents are needed.

**Level 2 — Dynamic Agents**: created per task from skill combinations like LEGO blocks. Live until task is complete, then destroyed.

**Level 3 — External Advisors**: GPT and Gemini connected via MCP servers. Not executors — advisors who give second opinions and participate in cross-model disputes.

## Installation

### Prerequisites

| Requirement | Required | How to get |
|------------|----------|------------|
| Claude Code | Yes | [claude.ai/code](https://claude.ai/code) — install CLI or desktop app |
| Claude subscription | Yes | Pro ($20/mo) or Max ($100-200/mo) |
| Node.js 18+ | Yes | [nodejs.org](https://nodejs.org/) |
| Git | Yes | [git-scm.com](https://git-scm.com/) |
| Codex CLI (GPT) | Optional | For GPT advisor in disputes |
| Gemini CLI | Optional | For Gemini advisor in disputes |

### Windows Installation

```powershell
# 1. Clone GODMODE
git clone https://github.com/neuron-one/GODMODE.git
cd GODMODE

# 2. Back up your existing config (if any)
if (Test-Path "$env:USERPROFILE\.claude\CLAUDE.md") {
    Copy-Item "$env:USERPROFILE\.claude\CLAUDE.md" "$env:USERPROFILE\.claude\CLAUDE.md.backup"
}

# 3. Copy GODMODE config
Copy-Item "claude\CLAUDE.md" "$env:USERPROFILE\.claude\CLAUDE.md"

# 4. Copy skills (create dir if needed)
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.claude\skills"
Copy-Item -Recurse -Force "skills\*" "$env:USERPROFILE\.claude\skills\"

# 5. Copy delegation rules
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.claude\rules\delegator"
Copy-Item -Force "rules\delegator\*" "$env:USERPROFILE\.claude\rules\delegator\"

# 6. Open Claude Code and install plugins
# Run these INSIDE Claude Code:
#   /plugin marketplace add jarrodwatts/claude-delegator
#   /plugin install claude-delegator@jarrodwatts-claude-delegator

# 7. (Optional) Install GPT advisor
npm install -g @openai/codex
# Then in Claude Code:
#   claude mcp add --transport stdio --scope user codex -- codex -m gpt-5.3-codex mcp-server
# Authenticate: run `codex login` in terminal

# 8. (Optional) Install Gemini advisor
npm install -g @google/gemini-cli
# Run `gemini` once to complete sign-in
# Then run /claude-delegator:setup in Claude Code to configure MCP

# 9. Restart Claude Code to load everything
```

### macOS / Linux Installation

```bash
# 1. Clone GODMODE
git clone https://github.com/neuron-one/GODMODE.git
cd GODMODE

# 2. Back up your existing config (if any)
[ -f ~/.claude/CLAUDE.md ] && cp ~/.claude/CLAUDE.md ~/.claude/CLAUDE.md.backup

# 3. Copy GODMODE config
cp claude/CLAUDE.md ~/.claude/CLAUDE.md

# 4. Copy skills
mkdir -p ~/.claude/skills
cp -r skills/* ~/.claude/skills/

# 5. Copy delegation rules
mkdir -p ~/.claude/rules/delegator
cp rules/delegator/* ~/.claude/rules/delegator/

# 6. Open Claude Code and install plugins
# Run these INSIDE Claude Code:
#   /plugin marketplace add jarrodwatts/claude-delegator
#   /plugin install claude-delegator@jarrodwatts-claude-delegator

# 7. (Optional) Install GPT advisor
npm install -g @openai/codex
# Then in Claude Code:
#   claude mcp add --transport stdio --scope user codex -- codex -m gpt-5.3-codex mcp-server
# Authenticate: run `codex login` in terminal

# 8. (Optional) Install Gemini advisor
npm install -g @google/gemini-cli
# Run `gemini` once to complete sign-in
# Then run /claude-delegator:setup in Claude Code to configure MCP

# 9. Restart Claude Code to load everything
```

### Verify Installation

After restarting Claude Code, check that everything works:

```
You: "What skills do you have?"
→ Claude should list 50+ skills across 11 categories

You: "Run a dispute on Redis vs PostgreSQL for task queues"
→ Claude should start a cross-model dispute (if GPT/Gemini configured)

You: "Review my code for security issues"
→ Claude should activate the code-security skill
```

### Troubleshooting

| Problem | Solution |
|---------|----------|
| Skills not showing | Check `~/.claude/skills/` has SKILL.md files |
| Plugins not working | Run `/reload-plugins` in Claude Code |
| GPT advisor fails | Run `codex login` in terminal |
| Gemini advisor fails | Run `gemini` once to authenticate |
| "Permission denied" on Mac | Run `chmod -R 755 ~/.claude/skills/` |

## Commands

After installation, these commands are available inside Claude Code:

### Core Commands

| Command | What it does |
|---------|-------------|
| `/interview [topic]` | Deep interview before starting a project (6 blocks: business, product, technical, security, launch, marketing) |
| `/dispute [topic]` | Cross-model dispute — Claude + GPT + Gemini debate until consensus |
| `/planning [task]` | Creates 3 tracking files (plan, findings, progress), eliminates drift |
| `/simplify` | 3 parallel agents review and clean code |
| `/cost-estimate` | Estimates project cost at market rates (shows ROI of AI) |
| `/self-improve` | Audits skills and CLAUDE.md, suggests improvements |
| `/codex:review` | GPT code review (read-only) |
| `/codex:adversarial-review` | GPT challenges your implementation decisions |
| `/codex:rescue` | Delegates a stuck problem to GPT |
| `/product-data-audit` | Full product ecosystem audit with interactive HTML report |

### Calling Advisors (natural language)

```
"Ask GPT what it thinks about..."        → GPT advisor
"Ask Gemini to evaluate this design..."   → Gemini advisor
"Run a dispute with GPT and Gemini on..." → Full cross-model dispute
```

### Keyboard Shortcuts

| Shortcut | Action |
|----------|--------|
| `Shift+Tab` | Switch modes (Plan / Code / Auto) |
| `Shift+Up` | Agent progress (Agent Teams) |
| `@file` | Add file to context |

## Skills Library

| Category | Skills | Focus |
|----------|--------|-------|
| **security** | 8 | OWASP audit, auth patterns, secrets scanning, supply chain, git guardrails |
| **development** | 7 | backend, frontend, database, Python, React, mobile |
| **infrastructure** | 7 | Docker, CI/CD, deploy, DNS/SSL, monitoring, server setup |
| **multimodel** | 7 | GPT advisor, Gemini advisor, adversarial review, LLM council |
| **design** | 5 | UI/UX pro max, components, mobile design, screen-to-code |
| **data** | 5 | scraping, data analysis, knowledge graphs |
| **workflow** | 5 | dispute, interview, planning, self-improve |
| **quality** | 4 | code review, batch operations, integration checks |
| **testing** | 4 | unit tests, E2E, browser testing, app store checks |
| **business** | 3 | cost estimation, marketing, product management |
| **ai-training** | 1 | ML model fine-tuning |

## Cross-Model Dispute Protocol

When you need to make an important technical decision:

```
Round 1: Claude + GPT + Gemini give independent positions
    |
Round 2: Each sees ALL others' positions and critiques
    |
Round 3: Defense + concessions
    |
Consensus? --NO--> Round 4-5: Negotiation
    |
   YES --> Final document with consensus/majority/unresolved
```

All three models participate as equals. No sycophantic agreement — adversarial roles are explicitly assigned.

## 16 Orchestrator Rules

1. **Plan First** — 3+ steps = plan mode, wait for approval
2. **Decompose** — 3+ files = subtasks, max 3-5 agents
3. **Verify** — no DONE without proof
4. **Self-Improve** — record lessons, propose CLAUDE.md updates (with approval)
5. **Security** — WARNING + safe alternative, never insecure patterns
6. **Post-Code** — list risks, test-first on bugs
7. **Permissions** — Auto Mode + deny list
8. **Memory** — hot/warm/cold hierarchy
9. **Output** — 64K tokens max
10. **Auto-Presets** — UI task = +ux skill, API task = +security
11. **Agent Communication Protocol** — JSON envelopes, validation, atomic writes
12. **Error Handling** — 3 classes: invalid spec, timeout, invalid output
13. **Structured Logging** — activity.jsonl for debugging
14. **Scratch Pad** — large outputs to file
15. **Compression** — save state at 70% context
16. **Simple Tasks** — no agents for simple work

## Ecomode (Smart Model Routing)

| Complexity | Model | Use For |
|-----------|-------|---------|
| Simple | Haiku/Sonnet | formatting, renaming, git ops |
| Medium | Sonnet | features, bug fixes, code review |
| Complex | Opus | architecture, security audit, disputes |

## Project Structure

```
GODMODE/
  claude/
    CLAUDE.md           --> Global orchestrator rules (copy to ~/.claude/)
  skills/
    security/           --> 8 security skills
    development/        --> 7 development skills
    design/             --> 5 design skills
    ...                 --> 50+ skills total
  rules/
    delegator/          --> GPT/Gemini delegation rules
  docs/
    ARCHITECTURE.md     --> Detailed architecture description
    INSTALL_GUIDE.md    --> Step-by-step installation
    SKILL_GUIDE.md      --> How to create custom skills
  examples/
    dispute-example.md  --> Real dispute transcript
    skill-example/      --> Example skill structure
```

## FAQ

**Q: Do I need GPT and Gemini?**
No. GODMODE works with Claude alone. GPT and Gemini are optional advisors for disputes and second opinions.

**Q: Is this a framework I need to code with?**
No. It's a configuration layer for Claude Code. You install it and it works. No programming required.

**Q: Can I add my own skills?**
Yes. See `docs/SKILL_GUIDE.md` for the skill format. Skills are just markdown files.

**Q: Will this conflict with my existing Claude Code setup?**
It adds to your `~/.claude/` directory. Back up your existing config before installing.

## Limitations

- This is a **showcase/research project**, not a supported product
- Built and tested on Windows 11 — Mac/Linux should work but not extensively tested
- Skills are opinionated — they reflect one developer's workflow
- External model integration requires separate subscriptions (ChatGPT, Gemini)
- Limited support — issues may not receive replies

## License

MIT License — use it however you want.

## Author

Built by [neuron-one](https://github.com/neuron-one) — a solo non-programmer who builds real products entirely with AI. No CS degree. No dev team. Just one person and an AI army.

*GODMODE is not affiliated with Anthropic, OpenAI, or Google.*

GODMODE — AI Development Team

Powered by NEURON.ONE Architecture

One person. Many projects. Corporate power in one pair of hands.

GODMODE turns Claude Code into a full AI development team. Multi-agent orchestration, cross-model disputes (Claude + GPT + Gemini), 50+ skills, smart model routing — all configured and ready to use.

Built by a non-programmer using AI. If I can do it, so can you.

What You Get

16 Orchestrator Rules — battle-tested operating principles for AI-assisted development
NEURON.ONE Architecture — adaptive swarm: one orchestrator + dynamic agents created per task
Cross-Model Disputes — Claude, GPT, and Gemini debate until consensus on architecture decisions
50+ Skills Library — security, development, design, testing, infrastructure, and more
Smart Model Routing (Ecomode) — Haiku for simple tasks, Sonnet for medium, Opus for complex
Agent Communication Protocol — JSON envelopes, preflight/postflight validation, atomic writes
Memory System — hot/warm/cold tiers with auto-persistence across conversations
External Advisors — GPT for security/architecture, Gemini for design/visual analysis

Architecture

GODMODE Architecture

Level 1 — Orchestrator: permanent Claude instance that understands the full picture, knows all skills, and decides how many agents are needed.

Level 2 — Dynamic Agents: created per task from skill combinations like LEGO blocks. Live until task is complete, then destroyed.

Level 3 — External Advisors: GPT and Gemini connected via MCP servers. Not executors — advisors who give second opinions and participate in cross-model disputes.

Installation

Prerequisites

Requirement	Required	How to get
Claude Code	Yes	claude.ai/code — install CLI or desktop app
Claude subscription	Yes	Pro ($20/mo) or Max ($100-200/mo)
Node.js 18+	Yes	nodejs.org
Git	Yes	git-scm.com
Codex CLI (GPT)	Optional	For GPT advisor in disputes
Gemini CLI	Optional	For Gemini advisor in disputes

Windows Installation

# 1. Clone GODMODE
git clone https://github.com/neuron-one/GODMODE.git
cd GODMODE

# 2. Back up your existing config (if any)
if (Test-Path "$env:USERPROFILE\.claude\CLAUDE.md") {
    Copy-Item "$env:USERPROFILE\.claude\CLAUDE.md" "$env:USERPROFILE\.claude\CLAUDE.md.backup"
}

# 3. Copy GODMODE config
Copy-Item "claude\CLAUDE.md" "$env:USERPROFILE\.claude\CLAUDE.md"

# 4. Copy skills (create dir if needed)
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.claude\skills"
Copy-Item -Recurse -Force "skills\*" "$env:USERPROFILE\.claude\skills\"

# 5. Copy delegation rules
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.claude\rules\delegator"
Copy-Item -Force "rules\delegator\*" "$env:USERPROFILE\.claude\rules\delegator\"

# 6. Open Claude Code and install plugins
# Run these INSIDE Claude Code:
#   /plugin marketplace add jarrodwatts/claude-delegator
#   /plugin install claude-delegator@jarrodwatts-claude-delegator

# 7. (Optional) Install GPT advisor
npm install -g @openai/codex
# Then in Claude Code:
#   claude mcp add --transport stdio --scope user codex -- codex -m gpt-5.3-codex mcp-server
# Authenticate: run `codex login` in terminal

# 8. (Optional) Install Gemini advisor
npm install -g @google/gemini-cli
# Run `gemini` once to complete sign-in
# Then run /claude-delegator:setup in Claude Code to configure MCP

# 9. Restart Claude Code to load everything

macOS / Linux Installation

# 1. Clone GODMODE
git clone https://github.com/neuron-one/GODMODE.git
cd GODMODE

# 2. Back up your existing config (if any)
[ -f ~/.claude/CLAUDE.md ] && cp ~/.claude/CLAUDE.md ~/.claude/CLAUDE.md.backup

# 3. Copy GODMODE config
cp claude/CLAUDE.md ~/.claude/CLAUDE.md

# 4. Copy skills
mkdir -p ~/.claude/skills
cp -r skills/* ~/.claude/skills/

# 5. Copy delegation rules
mkdir -p ~/.claude/rules/delegator
cp rules/delegator/* ~/.claude/rules/delegator/

# 6. Open Claude Code and install plugins
# Run these INSIDE Claude Code:
#   /plugin marketplace add jarrodwatts/claude-delegator
#   /plugin install claude-delegator@jarrodwatts-claude-delegator

# 7. (Optional) Install GPT advisor
npm install -g @openai/codex
# Then in Claude Code:
#   claude mcp add --transport stdio --scope user codex -- codex -m gpt-5.3-codex mcp-server
# Authenticate: run `codex login` in terminal

# 8. (Optional) Install Gemini advisor
npm install -g @google/gemini-cli
# Run `gemini` once to complete sign-in
# Then run /claude-delegator:setup in Claude Code to configure MCP

# 9. Restart Claude Code to load everything

Verify Installation

After restarting Claude Code, check that everything works:

You: "What skills do you have?"
→ Claude should list 50+ skills across 11 categories

You: "Run a dispute on Redis vs PostgreSQL for task queues"
→ Claude should start a cross-model dispute (if GPT/Gemini configured)

You: "Review my code for security issues"
→ Claude should activate the code-security skill

Troubleshooting

Problem	Solution
Skills not showing	Check `~/.claude/skills/` has SKILL.md files
Plugins not working	Run `/reload-plugins` in Claude Code
GPT advisor fails	Run `codex login` in terminal
Gemini advisor fails	Run `gemini` once to authenticate
"Permission denied" on Mac	Run `chmod -R 755 ~/.claude/skills/`

Commands

After installation, these commands are available inside Claude Code:

Core Commands

Command	What it does
`/interview [topic]`	Deep interview before starting a project (6 blocks: business, product, technical, security, launch, marketing)
`/dispute [topic]`	Cross-model dispute — Claude + GPT + Gemini debate until consensus
`/planning [task]`	Creates 3 tracking files (plan, findings, progress), eliminates drift
`/simplify`	3 parallel agents review and clean code
`/cost-estimate`	Estimates project cost at market rates (shows ROI of AI)
`/self-improve`	Audits skills and CLAUDE.md, suggests improvements
`/codex:review`	GPT code review (read-only)
`/codex:adversarial-review`	GPT challenges your implementation decisions
`/codex:rescue`	Delegates a stuck problem to GPT
`/product-data-audit`	Full product ecosystem audit with interactive HTML report

Calling Advisors (natural language)

"Ask GPT what it thinks about..."        → GPT advisor
"Ask Gemini to evaluate this design..."   → Gemini advisor
"Run a dispute with GPT and Gemini on..." → Full cross-model dispute

Keyboard Shortcuts

Shortcut	Action
`Shift+Tab`	Switch modes (Plan / Code / Auto)
`Shift+Up`	Agent progress (Agent Teams)
`@file`	Add file to context

Skills Library

Category	Skills	Focus
security	8	OWASP audit, auth patterns, secrets scanning, supply chain, git guardrails
development	7	backend, frontend, database, Python, React, mobile
infrastructure	7	Docker, CI/CD, deploy, DNS/SSL, monitoring, server setup
multimodel	7	GPT advisor, Gemini advisor, adversarial review, LLM council
design	5	UI/UX pro max, components, mobile design, screen-to-code
data	5	scraping, data analysis, knowledge graphs
workflow	5	dispute, interview, planning, self-improve
quality	4	code review, batch operations, integration checks
testing	4	unit tests, E2E, browser testing, app store checks
business	3	cost estimation, marketing, product management
ai-training	1	ML model fine-tuning

Cross-Model Dispute Protocol

When you need to make an important technical decision:

Round 1: Claude + GPT + Gemini give independent positions
    |
Round 2: Each sees ALL others' positions and critiques
    |
Round 3: Defense + concessions
    |
Consensus? --NO--> Round 4-5: Negotiation
    |
   YES --> Final document with consensus/majority/unresolved

All three models participate as equals. No sycophantic agreement — adversarial roles are explicitly assigned.

16 Orchestrator Rules

Plan First — 3+ steps = plan mode, wait for approval
Decompose — 3+ files = subtasks, max 3-5 agents
Verify — no DONE without proof
Self-Improve — record lessons, propose CLAUDE.md updates (with approval)
Security — WARNING + safe alternative, never insecure patterns
Post-Code — list risks, test-first on bugs
Permissions — Auto Mode + deny list
Memory — hot/warm/cold hierarchy
Output — 64K tokens max
Auto-Presets — UI task = +ux skill, API task = +security
Agent Communication Protocol — JSON envelopes, validation, atomic writes
Error Handling — 3 classes: invalid spec, timeout, invalid output
Structured Logging — activity.jsonl for debugging
Scratch Pad — large outputs to file
Compression — save state at 70% context
Simple Tasks — no agents for simple work

Ecomode (Smart Model Routing)

Complexity	Model	Use For
Simple	Haiku/Sonnet	formatting, renaming, git ops
Medium	Sonnet	features, bug fixes, code review
Complex	Opus	architecture, security audit, disputes

Project Structure

GODMODE/
  claude/
    CLAUDE.md           --> Global orchestrator rules (copy to ~/.claude/)
  skills/
    security/           --> 8 security skills
    development/        --> 7 development skills
    design/             --> 5 design skills
    ...                 --> 50+ skills total
  rules/
    delegator/          --> GPT/Gemini delegation rules
  docs/
    ARCHITECTURE.md     --> Detailed architecture description
    INSTALL_GUIDE.md    --> Step-by-step installation
    SKILL_GUIDE.md      --> How to create custom skills
  examples/
    dispute-example.md  --> Real dispute transcript
    skill-example/      --> Example skill structure

FAQ

Q: Do I need GPT and Gemini? No. GODMODE works with Claude alone. GPT and Gemini are optional advisors for disputes and second opinions.

Q: Is this a framework I need to code with? No. It's a configuration layer for Claude Code. You install it and it works. No programming required.

Q: Can I add my own skills? Yes. See docs/SKILL_GUIDE.md for the skill format. Skills are just markdown files.

Q: Will this conflict with my existing Claude Code setup? It adds to your ~/.claude/ directory. Back up your existing config before installing.

Limitations

This is a showcase/research project, not a supported product
Built and tested on Windows 11 — Mac/Linux should work but not extensively tested
Skills are opinionated — they reflect one developer's workflow
External model integration requires separate subscriptions (ChatGPT, Gemini)
Limited support — issues may not receive replies

License

MIT License — use it however you want.

Author

Built by neuron-one — a solo non-programmer who builds real products entirely with AI. No CS degree. No dev team. Just one person and an AI army.

GODMODE is not affiliated with Anthropic, OpenAI, or Google.