cc-godmode

---

The System That Builds Itself

Welcome to the machine shop. Except the machines are building themselves — and they're getting better at it.

CC_GodMode v7.0.0 and v7.1.0 weren't built by hand. They were planned, architected, implemented, validated, documented, and shipped by the exact agents defined in this repo. PRs #22 and #23 are proof. The orchestrator delegated to @architect for design, @builder for implementation, @validator and @tester for quality gates, and @scribe for documentation. Then it created the PR itself.

That's not marketing copy. That's what happened.

---

Install in 30 Seconds

CC_GodMode is a Claude Code plugin that transforms your development workflow into a self-orchestrating multi-agent system. Here's how to set it up:

macOS / Linux:

git clone https://github.com/cubetribe/ClaudeCode_GodMode-On.git
cd ClaudeCode_GodMode-On
./scripts/apply-global-claude-setup.sh
./scripts/apply-global-claude-setup.sh --check   # Verify installation

Windows (PowerShell):

git clone https://github.com/cubetribe/ClaudeCode_GodMode-On.git
cd ClaudeCode_GodMode-On
.\scripts\apply-global-claude-setup.ps1
.\scripts\apply-global-claude-setup.ps1 -Check   # Verify installation

Then activate in any project:

cd your-project
cp ~/.claude/templates/CLAUDE-ORCHESTRATOR.md ./CLAUDE.md
claude

Done. The orchestrator is active.

---

Just Type GodMode

Once installed, daily usage is one line:

GodMode: New Feature: user authentication with JWT
GodMode: Bug Fix: cart total miscalculates with discount codes
GodMode: Research: best approach for real-time sync in React 18

No installation ceremony. After one-time setup, most sessions need nothing — the orchestrator loads from CLAUDE.md automatically. Use the ContextRestore prompt only after /compact or if the orchestrator loses its role. You say what you want — the system figures out which agents to call, in what order, at what cost.

---

What Changed in v7 — and Why

CC_GodMode v7.0.0 was redesigned based on how Claude Fable 5 actually works, not how earlier models needed to be prompted:

• Persona prompts are gone. "You are a Senior X with 15 years of experience" adds nothing on Fable 5 — identity is now one line, the task description does the work.
• Aggressive caps removed. Extensive CRITICAL/MUST/NEVER scaffolding in older agents caused overtriggering. Per Anthropic's guidance, normal imperative phrasing is enough; hard safety rules (never push, api-guardian mandatory) keep their weight.
• Effort fields instead of verbose instructions. Each agent carries an effort: frontmatter field (Claude Code ≥2.1.152) that tunes token budgets automatically — no long preambles needed.
• Smart Routing is now the default. The old system always ran every agent. v7 runs only what the risk level requires — estimated 30–50% token reduction for standard features.
• Verdict contract instead of report re-reading. Agents return a structured 3-bullet STATUS verdict to the orchestrator. The full on-disk report exists for humans and for BLOCKED escalations.
• Plugin-based install replaces copy-paste prompts as the primary path. .claude-plugin/plugin.json is the modern distribution mechanism; the install prompts stay as a manual fallback.

---

The Story

It started simple: One developer, mass sleep deprivation, and a vision.

Phase 1: Manual labor. Researching best practices. Reading docs. Testing prompts. Failing. Iterating. Building agent after agent. Workflow after workflow. Week after week.

Phase 2: The system works. 8 specialized AI agents orchestrating themselves. Features get built. Bugs get fixed. Documentation writes itself. "This is pretty good," I thought.

Phase 3: January 6th, 2026. A thought: "What if I use the system... to improve the system?"

I gave it one prompt. The orchestrator delegated to the research team. Analyzed its own architecture. Found inefficiencies. Proposed improvements. Implemented them. Validated itself. Documented the changes.

The loop closed.