Claude Framework — Development skill for Claude Code

Claude Framework

**A working method for Claude Code, installed in one command and kept honest by a doctor.**

Specialised subagents, delegation rules, evidence-before-action and a context budget per role — generated into your project, versioned, and checked by a tool that tells you the moment an installation drifts.

claude-framework-eng/
├── method/
├── coordinator/
├── cycles/
├── agents/
├── profiles/
├── shared/
│   ├── core/
│   └── domain/
├── hooks/
├── output-styles/
├── templates/
├── skills/
└── tools/
    └── fwbuild/


---

Why

Claude Code gives you subagents, hooks and skills. It does not give you a **method** — so everyone writes one by hand, per project, and watches it drift.

• Agents that stay in their lane. A roster picked for your project, a routing table, one task per agent with a verifiable done-criterion, and one fixed report every subagent closes with: confidence, what would disprove it, what it did not verify.
• Context you actually use. The method every agent needs lives in CLAUDE.md; delegation lives where only the coordinator reads it; guides load on demand. fwbuild cost turns the words into tokens and dollars.
• Guardrails that don't trust the model. Hooks block --no-verify and edits that loosen an existing linter config. Read-only reviewers get no shell at all.
• One method, every project. The generated method sits in a hashed kernel region: edit it and the doctor sees it, and framework-sync carries the edits worth keeping back into the source.

---

Install

Once per machine. Requires Claude Code and Python 3.11+ — stdlib only, nothing to install. Two self-standing editions: swap `claude-framework-eng` for `claude-framework-it` to work in Italian.

git clone https://github.com/290990-collab/Claude-Framework.git ~/.claude/claude-framework