My daily Claude Code setup — what actually works

Claude Code has become my main work tool over the past few months. Not just for generating code — for everything: codebase exploration, debugging, agent prototyping, spec writing.

Here’s what actually works in my setup, and what has disappointed me.

What I actually use

For exploring an unknown codebase

The first thing I do on a new client project: launch Claude Code and ask it to explain the architecture. Not “generate code for me,” but “explain how this module works and what the critical dependencies are.”

The efficiency is remarkable. In 20 minutes I have an overview that would have taken me 2 hours to build alone by reading the code.

For prototyping system prompts

When I need to design a system prompt for an agent, I always start by discussing it with Claude Code. I describe the expected behavior, it proposes a structure, I critique, we iterate.

It doesn’t write the final prompt — I do. But the conversation helps me clarify my ideas much faster than working alone.

For spec reviews

Before validating a spec with a dev team, I run it through Claude Code with a “skeptical senior developer” role. The question: “What’s missing in this spec? What will cause problems at implementation?”

The result is often surprising — it identifies ambiguities I had missed.

My CLAUDE.md configuration

The CLAUDE.md file at the project root is the key. Here’s what I always put in it:

# Project context
[Project description, tech stack, constraints]

# Conventions
- Language: English for comments and documentation
- Style: [your code conventions]

# What you must NOT do
- Don't modify configuration files without validation
- Don't touch database migrations
- Always ask for confirmation before deleting code

# Business context
[Glossary of domain-specific terms]

What disappointed me

Long contexts. Beyond a certain codebase size, performance degrades. I now segment my sessions by module.

Test generation. Automatically generated tests are often too superficial — they test the happy path without really challenging edge cases.

The overengineering tendency. Without explicit simplicity instruction, Claude Code may produce a generic extensible solution where you wanted something simple. “Prefer the simplest solution” in CLAUDE.md helps a lot.