AI Integration — Claude in Production

Introduction

I use Claude Code as my primary tool for my work—development, writing, and automation. The goal is to establish a structured, reproducible methodology that can be applied to any technical environment. This project documents how I configure, use, and monitor AI in my projects.

Why structure your use of AI

Most people use AI in a one-off question-and-answer mode. This yields mediocre and inconsistent results. What makes all the difference is having a method: knowing how to break down a problem for the AI, providing it with the right context, maintaining a project history, and automating whatever can be automated—all while retaining control over every action. Without that, AI produces code we don’t understand, files we can’t track, and results we can’t reproduce. And worse still, our brains stop working altogether.

How it works in practice

Every Claude Code project starts with a CLAUDE.md file that defines the AI’s behavior: coding conventions, tech stack, business rules, and links to my Obsidian knowledge base for broader context. When I’m developing, Claude Code has access to the entire codebase within the IDE. I approve every interaction—there’s no standalone mode. I see the suggested changes in real time, accept or reject them, and maintain control over versioning. The result remains consistent from one session to the next.

Three levels of use

Daily Use — Development, Writing, Debugging — Claude Code is always open in the terminal. I use it to navigate the codebase, edit files, run commands, write content, and debug. Every action triggers a validation and adds information to its local structure. Project configuration — CLAUDE.md, memory, Obsidian — Each project has its own CLAUDE.md file that defines the context: stack, conventions, folder structure, and links to relevant Obsidian notes. The persistent memory accumulates feedback from one session to the next; the organization within Obsidian is crucial to avoid overwhelming the context with the entire conversation. Hooks trigger automatic checks. Technical integration — API, MCP, custom skills — Integrating the Anthropic API into backend applications. Configuring MCP servers to connect Claude to external tools (Notion, Airtable, databases). Creating reusable skills to automate recurring workflows: generating formatted content, code review, deployment. Caution regarding MCPs: (disclaimer: danger—use only the official version).

Working Principles

Constant human oversight — Every change suggested by Claude is visible in the IDE before it’s applied. No “YOLO” mode. I approve every interaction and review every diff. Systematic code review — I use the review features built into Claude Code to scan the generated code: detecting security vulnerabilities, logical errors, and inconsistencies with project conventions. The AI review complements my manual review; it does not replace it. Cleanup of generated files — Claude Code generates configuration, memory, and cache files. I systematically check what is created, limit the files that are committed, and maintain a strict .gitignore file to ensure nothing related to the AI pollutes the repository. The origin of every file in the repo is traceable. Local Execution — Everything runs locally. The code stays on my machine, commands run in my terminal, and files never leave my environment. The only remote calls are API requests to Claude—the rest is under my control. Reproducible in any context — The method doesn’t depend on a specific project or stack. The same setup (CLAUDE.md, memory, hooks, skills) can be applied to a Next.js project, a Python backend, a Terraform repo, or a data pipeline. What changes is the content of the context file—not the method.

Technology stack

AI — Claude (Opus, Sonnet, Haiku), Anthropic API, Claude CLI & IDE Configuration — CLAUDE.md, persistent project state, hooks, custom skills Knowledge Base — Obsidian (linked notes, extended project context) Integration — MCP servers (Notion, Airtable, etc.), Anthropic SDK (Python/TypeScript) Security — Automated code review, strict .gitignore, run locally whenever possible Environment — Terminal, VS Code, Git