sangrokjung/claude-forge

oh-my-zsh for Claude Code — 16 agents, 35 commands, 32 skills, 21 safety hooks in one install. v4.0 adds an adversarial review loop: a second agent that never sees the first one's reasoning. MIT.

What it solves

Claude Code is a powerful terminal-based AI coding assistant, but it lacks built-in organizational procedures, safety guardrails, and specialized roles. Claude Forge provides a comprehensive "equipment pack" that transforms the bare assistant into a professional-grade development environment by adding structured workflows, safety checks, and a library of specialist agents.

How it works

Claude Forge extends Claude Code through a combination of configuration files and plugins. It implements a multi-layered system consisting of:

  • Specialist Agents: 16 dedicated AI roles (e.g., planner, security reviewer, TDD guide) that Claude can delegate tasks to.
  • Slash Commands: 35 shortcuts (like /plan or /tdd) that trigger complex, pre-defined workflows.
  • Skills: 33 saved procedures or "playbooks" that the AI follows automatically for specific tasks.
  • Hooks: 21 safety checks that run silently in the background to block dangerous actions, such as leaking secrets or executing destructive database commands.
  • Rules: 14 behavioral guidelines that define coding styles and security principles.
  • MCP Connections: Integration with external tools like browser automation (Playwright) and live documentation search.

Who it’s for

Developers using Claude Code who want to automate their development pipeline (plan $\rightarrow$ test $\rightarrow$ review $\rightarrow$ verify $\rightarrow$ ship) and add enterprise-level safety and reliability to their AI-assisted coding.

Highlights

  • Adversarial Verification Loop: A "maker $\neq$ checker" system where an independent agent verifies changes without seeing the original agent's reasoning.
  • 6-Layer Security Defense: Built-in hooks that automatically block leaked passwords and unsafe remote scripts.
  • Reliability Pack: Features like API-error auto-resume and doom-loop guards to prevent session crashes.
  • Loop-Forge: A tool to turn repetitive tasks into reusable, self-guarding slash commands.
  • Harness-Diet: A utility to measure and shrink always-loaded context to stay under token budgets without losing governance.

関連

  • プロジェクト
  • プロジェクト
  • プロジェクト
  • プロジェクト
  • プロジェクト