An engineering framework for AI coding agents.
Stop hoping your agent writes good code — make it.
Quick Start
claude plugin marketplace add NoobyGains/godmode
claude plugin install godmode
That's it. mention Godmode — GodMode activates automatically.
The Problem
AI coding agents have a predictable failure mode: you ask for a feature, and they immediately start writing code. No questions about requirements. No check for existing patterns. No tests. No spec. Just a wall of generated files and a "let me know if you need changes."
The result is code that works in isolation but clashes with everything around it. Inconsistent naming. Duplicated logic. Tests bolted on after the fact. Entire components built from scratch when a solid open-source solution already exists. You end up spending more time fixing the agent's output than you would have spent writing it yourself.
GodMode fixes this. It is a 36-skill system that activates automatically and governs the full development lifecycle — from the first question about requirements through to a merged pull request.
How It Works
GodMode operates as a five-phase gated pipeline. The agent cannot advance until each phase meets its exit criteria.
DEFINE — The agent interrogates the problem space. It asks targeted questions one at a time, surfaces trade-offs, researches proven references, presents 2-3 approaches with rationale, and produces a spec document. Nothing moves forward until you approve it.
PLAN — The approved spec is decomposed into atomic tasks (2-5 minutes each). Every task includes exact file paths, complete code snippets, shell commands with expected output, and pass/fail criteria. Ambiguity dies here.
EXECUTE — Fresh subagents receive individual tasks and work in isolation. Test-driven development is mandatory: write a failing test, make it pass, refactor, commit. A two-stage review (spec compliance + code quality) runs after every task.
REVIEW — Completed work is evaluated against the original spec, then against code quality standards. Issues are severity-ranked. Critical findings block forward progress until resolved.
SHIP — End-to-end test verification. The agent presents options: merge, open a PR, keep the branch, or discard. Worktrees are cleaned up. Nothing left dangling.
Dynamic Execution Routing
GodMode doesn't just have one way of working. It analyzes your task and automatically selects the optimal execution mode:
Task arrives
│
├── Single atomic step?
│ └── YES → Handle directly (no orchestration)
│
├── Multiple independent parts?
│ ├── Need peer coordination? → Agent Teams
│ └── No coordination needed? → Parallel Execution
│
└── Multi-step sequence?
└── YES → Delegated Execution (sequential + review gates)
Four Execution Modes
Direct — Simple tasks that don't need decomposition. Config changes, typo fixes, single-file modifications.
Delegated Execution — Sequential multi-step implementation. A fresh subagent handles each task, followed by a spec compliance audit and a code quality audit. The controller never loses context because it curates exactly what each subagent needs.
Parallel Execution — Independent problems dispatched to concurrent agents. Each agent works in isolation with a focused brief. Results are reconciled and integration-tested after all agents report back. Used for multi-domain debugging, distributed investigation, independent test fixes.
Agent Teams — Multiple Claude instances collaborating via peer-to-peer messaging, each in its own git worktree. Teammates can share discoveries mid-work, negotiate interfaces in real-time, and coordinate on shared problems. Used for multi-module features, codebase migrations, and diagnosis of complex issues.
The system decides automatically. You say "build me a dashboard" and GodMode figures out whether that's a 3-agent team job or a sequential delegation.