# How I Built a Production Claude Code Setup (and Open-Sourced It) > Source: > Published: 2026-06-22 07:06:51+00:00 I have been using Claude Code seriously for months. And I kept hitting the same frustrations: So I built a reference architecture that fixes these. Then I open-sourced it. [Claude Code Blueprint](https://github.com/faizkhairi/claude-code-blueprint) is a library of ready-to-copy files: CLAUDE.md, hooks, agents, skills, and rules. You copy what you need into your own project. Nothing more. It is not a framework. It is not a wrapper. It is configuration files that change how Claude Code behaves. **12 agents * 17 skills * 12 hooks * 6 rules** 60 seconds to start: `ash` curl -o CLAUDE.md https://raw.githubusercontent.com/faizkhairi/claude-code-blueprint/main/CLAUDE.md The CLAUDE.md ships with three behavioral rules that prevent the most common AI coding mistakes: Before writing any fix, Claude must run four checks: git state, error source, existing suppression, minimum viable diagnosis. This stops the most common waste -- building an elaborate fix for a problem that was already solved in a previous commit. Any change touching more than one file goes through plan mode. No exceptions. Claude proposes, you approve, then it executes. This prevents the pattern where Claude charges into a 15-file refactor and produces something you did not ask for. After finishing any task, Claude runs a verification pass appropriate to the work type. API endpoint? It curls the live URL. Config change? It confirms the value was actually picked up. File edit? It re-reads the changed block. This alone catches a significant proportion of errors before you do. This was the hardest part to get right. Claude Code has no native persistent memory between sessions. The blueprint solves this with a file-based memory system: Every session starts with a load-session hook that reads these files. Every session ends with a save-session hook that writes them. The result: Claude picks up where it left off, even days later. One of the things I learned the hard way: running everything on the expensive model wastes quota on mechanical work. The blueprint ships with 12 agents, each pinned to the right model tier: The main Claude Code session (Opus) orchestrates. Sonnet subagents do the bulk work. Your expensive quota goes to the decisions that need it. Skills are slash commands that invoke multi-step workflows. The blueprint ships 17: | Skill | What It Does | |---|---| | eview-full | Full PR review across security, logic, tests, style | | deploy-check | Pre-deploy checklist for your stack | | est-check | Run tests, check coverage, flag gaps | | sprint-plan | Turn a backlog into a sprint with estimates | | load-session | Restore context from memory at session start | | save-session | Write session state to memory at session end | | self-review | Review your own diff before pushing | Hooks run outside Claude's context window. They cost zero tokens. They are the highest-leverage place to add guardrails. The blueprint ships 12: Any developer using Claude Code who wants more reliable, consistent behavior. The blueprint is framework-agnostic -- it configures Claude's behavior, not your code. There are four presets if you want a guided start: `ash` ./setup.sh --preset=minimal # CLAUDE.md + 3 core hooks ./setup.sh --preset=standard # + agents + skills ./setup.sh --preset=full # everything The README and setup docs are available in Japanese, Korean, and Chinese. The community contributed these. GitHub: [github.com/faizkhairi/claude-code-blueprint](https://github.com/faizkhairi/claude-code-blueprint) MIT licensed. Copy what you need. Leave the rest. If it is useful, a star helps others find it.