Phasecraft — Development skill for Claude Code

PhaseCraft — An Opinionated Agentic Development Framework

PhaseCraft is a multi-agent SDLC framework for [Claude Code](https://docs.anthropic.com/en/docs/claude-code). It implements a lifecycle-staged, spec-driven approach to software development with AI agents. You get a team of specialized AI agents — Product Analyst, UI Specialist, Architect, Coder, QA Engineer and Project Manager, each with well defined role boundaries and scoped file access. Every piece of work has a clear role owner. You run slash commands; agents handle spec, design, architecture, implementation, and review through a structured lifecycle.

---

**Treating AI as (engineer + PM + architect + domain-expert) all-in-one leads to drift, churn, and degrading code over time.**

PhaseCraft separates these roles into bounded agents operating across distinct lifecycle stages, with humans-in-the-loop as checkpoint gates.

---

For the thinking behind this approach, see [The Lure of Malleability in AI-generated Software](https://adaptive.svbtle.com/the-lure-of-malleability-in-ai-generated-software).

Work is organized into phases — coherent, dependency-ordered sets of features, each fully built, tested, and reviewed before the next phase begins. Think of a **phase as a sprint scoped for AI**: bounded by token limits and context drift, not time.

A *session* in this README means one Claude Code conversation — from the moment you type a slash command to when the agent prints its handoff summary. **Important:** Each slash command should run in its own session to avoid context bleed between agents.


**Beta** — this framework is under active development and testing on Claude Code via Claude Desktop. Other environments have not been tested. Expect rough edges, submit feedback/bugs and check back for updates.

---

How this README is structured

**