superforge-dev

# Superforge Dev — Multi-Agent Building & Model Tiering Engine Use this skill when implementing multi-component features, executing complex build plans, or dispatching subagents. It ensures optimal model tiering and agent topology selection. --- ## 1. Agent Topology Proposal (Subagents vs Agent Teams) Before dispatching, evaluate and **explicitly notify the user** of the recommended structure: - **Subagents Pattern (Default — Low Token Cost)**: Use for isolated, modular tasks (building components, fixing bugs, writing tests). One-way dispatch. - **Agent Teams Pattern (Interactive — High Token Cost)**: Use when cross-perspective debate is required (architecture trade-offs, multi-agent debate). *Notification Template*: > *"Proposing **Subagents Pattern** (Sonnet 5 workers) for fast, token-efficient execution. Say 'use Agent Teams' if you prefer interactive multi-agent debate."* --- ## 2. Model Tier Classification (Claude 5 & Multi-LLM) ### Golden Rule for Claude 5: > **"判断は Opus 5, 量は Sonnet 5, 雑務は Haiku 4.5, 持久戦は Fable 5"** | Subtask Nature | Claude 5 | Gemini Env | Codex Env | Kimi Env | Effort | |---|---|---|---|---|---| | **Tier A (Judgment & Architecture)**: System architecture, PRD, code review, security | `Opus 5` | `3.6-flash` | `Codex Sol` | `K3 Max` | `high`/`xhigh` | | **Tier A (Unattended Long Runs)**: Overnight builds, 10+ step refactoring | `Fable 5` | `3.6-flash` | `Codex Sol` | `K3 Max` | `high` | | **Tier B (Volume Implementation)**: Feature workers, ...