Claude Code in Practical Use

Of the agentic tools, Claude Code is our daily driver. It works in the terminal and in the IDE, reads the project, edits files, runs builds and tests — all while adhering to project-specific rules.

What makes it strong in daily use

• Project context: A CLAUDE.md in the repo gives the agent the house rules — conventions, deployment steps, taboos. It reads them at every session and sticks to them, without you having to explain it all again.
• Skills: Recurring tasks (security audit, SEO check, integration check) we encapsulate as Skills — retrievable specialist knowledge instead of copy-paste prompts that go stale and drift.
• MCP server: Via the Model Context Protocol, the agent gets tools put in its hands — e.g., current documentation, browser automation, or custom data sources. So it relies on the real current state, not on guessed knowledge.
• Subagents: Large tasks are split and processed in parallel, rather than feeding one endless session.

How Skills, MCP, and Subagents Work Together

The three mechanisms complement each other. A typical run looks like this: the main agent receives a goal, pulls in a skill when needed (say, the security audit), fetches live data via an MCP server (the current library version, the live inventory), and delegates heavy research to a subagent. The result comes back structured, and the main agent decides the next step.

The key point: these tools are not an end in themselves. We install them where they measurably help — and otherwise leave it at a clear prompt. Outfitting everything with skills and MCP servers builds you a second, invisible system that wants to be maintained.

How we use it

From refactoring over bug fixing to the complete setup of applications — this website itself was created with Claude Code, including deployment behind our reverse proxy. This very blog post was co-produced: the agent reads the inventory, proposes expansions, runs the edits — and we read everything before it is committed.

Where Claude Code is strongest: in large, unfamiliar codebases, when scaffolding recurring structures, and for routine refactoring. Where it is weaker: on hard architecture decisions and on anything that demands real domain knowledge.

Whether It Pays Off for the SME

The question is not whether Claude Code works, but whether it pays off in concrete operation. For us the answer is yes — but with three conditions. First: someone has to read the output critically, or the codebase quickly gets worse instead of better. Second: the project rules in CLAUDE.md must be maintained, or the agent relies on outdated assumptions. Third: sensitive steps stay on a short leash; production and secrets are never delegated blindly.

Hold to those three and you get an assistant that clears routine in minutes and takes load off senior developers. Ignore them and you quickly get a codebase no one understands anymore.

Taking the limits seriously

• Review remains mandatory — generated code is read, not blindly adopted.
• Verification before "done" — run it, observe, then believe.
• Sensitive actions (production, secrets) belong on a short leash.

Claude Code replaces neither an engineer nor an engineering professional. It makes good faster — and unfortunately also bad habits. The difference lies in the craftsmanship around it.