Claude Framework
Autonomous development framework turning Claude Code into a full engineering team
The Problem
Working with Claude Code as a solo developer is powerful, but without structure it devolves into ad-hoc prompting. Every session starts cold — no memory of past decisions, no consistent review process, no way to enforce quality gates. You end up re-explaining project context, manually running the same review checklists, and losing institutional knowledge between sessions.
The core problem is operational: how do you make an AI coding assistant behave like a full engineering team — with specialized reviewers, automated workflows, persistent memory, and asynchronous communication — when you're a team of one? Existing tooling didn't address this. Claude Code's skill and hook systems provide primitives, but there was no opinionated framework for composing them into a repeatable development workflow.
The Approach
The framework is structured as a global ~/.claude/ directory that layers reusable skills, specialized agents, lifecycle hooks, and a communication pipeline on top of Claude Code's primitives. Every project inherits the global framework and can override behavior through per-project config files — JSON over code, always.
Skills use YAML frontmatter for metadata: context: fork for isolation,allowed-tools for least-privilege access, anddisable-model-invocation to prevent accidental cost multiplication. Review skills run in forked contexts so they can't pollute the main session. Orchestrator skills like /milestone-review and /gha-checkup spawn multiple skills in parallel.
The communication layer routes Claude's built-in hooks (approvals, questions, notifications) through a unified dispatcher to Slack or Discord, selected per-project via config. Session notes flow into an Obsidian vault, building a searchable knowledge graph of every development session across all projects.
Key Features
Skill System with Frontmatter
35+ global skills with YAML frontmatter controlling context isolation, tool restrictions, and auto-invocation. Skills range from review agents and deployment automation to game design workflows and session management. Orchestrator skills compose multiple skills in parallel.
Agent Orchestration
Seven specialized review agents (security, architecture, product, UI design, business strategy, QA) evaluate projects at milestones. Each agent uses project-specific config from .claude/review-config.json. A deploy agent handles the full lifecycle from GitHub repo to production.
Communication Pipelines
Unified dispatcher routes Claude Code hooks to Slack or Discord per-project. Supports approval gates with timeout polling, question threads, fire-and-forget notifications, and feedback channels with emoji-based triage. A message classifier and routing engine handle inbound items from mobile.
Obsidian Knowledge Graph
Session notes, daily wraps, and decision logs flow into an Obsidian vault organized by date. The /start-day skill reads yesterday's notes for continuity. The /wrap skill captures session summaries. Over time this builds a searchable second brain of every architectural decision and debugging session.
Technical Deep Dive
Results & Learnings
The framework now manages 10+ active projects — from SaaS platforms and crypto trading tools to PWA games and homelab infrastructure. It has grown from 5 review agents to 35+ skills across development, deployment, design, operations, and session management, all built incrementally over 10 days of active development.
- Context isolation via
context: forkwas the single most important design decision — without it, review skills would pollute the working session and orchestrator skills would multiply costs unpredictably - Config over code proved essential for scaling across projects — adding a new project to the framework is a JSON file, not a code change
- The communication layer transformed Claude from a local tool into an asynchronous collaborator — approval gates and feedback channels mean work continues while away from the terminal
- Building meta-tooling (skills that create skills, reviews that review the framework) created a compounding effect — the framework improves itself with each session