Claude Code

Claude Code is an AI coding agent by Anthropic that works directly in the repository: it reads and searches files, runs commands, and edits code. Claude is available company-wide at Kiwee, and most of our developers now use it daily. Unlike autocomplete tools that suggest the next lines of code, Claude Code works as an agent: you describe a task, and it explores the codebase and makes the changes itself.

How we have used it so far

• Navigating the codebase. Claude Code quickly finds where something is implemented, even in large projects. In Magento, for example, it can search both our project files and the vendor code to tell whether a behavior comes from our code, a third-party module, or core. In our setup, the vendor code lives only inside the Docker container, which makes searching it by hand slow and inconvenient. Claude Code simply runs the search there with docker compose exec, with no extra setup needed.
• Understanding code. We can ask questions about unfamiliar code and get answers fast. This is especially useful when jumping into a legacy project with little or no documentation: Claude Code can go through the git history and commit messages to explain why something was built a certain way, and with Jira connected through an MCP server, it can even read the tickets behind old branches. It feels like talking to someone who has already worked on the project.
• Writing code and tests. It speeds up routine implementation and writing tests, leaving us more time for the parts that need careful thought.
• Repetitive checks and changes. It handles tedious work at scale, for example going through 115 email templates stored in the database, finding the ones with a formatting mistake, and fixing them, which would have taken hours manually.
• Self-review before code review. Typos have practically disappeared from our pull requests, and it spots related code or documentation that should change too.

Claude Code can also connect directly to Claude in Chrome, a browser extension, and control a real Chrome window. We use it to verify fixes and features right after Claude Code has implemented them. It goes through the flow the way a customer would, for example placing an order, and can inspect the page source, console logs, and network requests directly in the browser. Claude Code sees the result of its own change right away, which makes debugging and verification much faster.

CLAUDE.md

Claude Code works better when it knows the project. For that, it reads a CLAUDE.md file at the start of each session. Ours contains conventions, a project description, and known problems with their solutions. The whole team keeps it up to date, adding new modules or solutions as soon as they come up. With a good CLAUDE.md, answers get faster, especially for problems we have seen before. At the same time, outdated notes can do harm: Claude trusts what is written there, so an old entry may lead it in the wrong direction.

What to watch out for

Outdated notes are not the only thing to keep an eye on:

• Quick fixes over maintainable ones. Claude sometimes makes bigger changes than needed. For example, to change a text color, it tried to duplicate a whole template file instead of adding one line of CSS, which would have left us with extra code to maintain. Asking for a plan first gives you a chance to adjust the approach or add your own ideas before any code is written. Plan Mode is made for exactly that.
• Overconfidence. Claude tends to stick with the first plausible explanation and present it as a fact. Be careful with words like "only", "never" or "definitely", and ask for evidence: a command, its output, or a line of code.
• Knowledge cutoff. Claude can search the web, but does not always realize that it should. If an answer seems outdated or something cannot be found, the reason may be that the model simply does not know about it yet. In that case, point Claude in the right direction: share a link or the documentation, or ask it to search.
• Model choice. Results depend a lot on the model. When we started using Claude Code at the beginning of 2026, the default Opus 4.6 worked well, but after switching to Opus 5, replies often got unnecessarily long, with the actual answer buried somewhere inside. The recently released Opus 5.5 has been much better in this respect, so if the results suddenly get worse, switching the model is worth a try.
• Effort level. More effort is not always better. On a simple task, a high effort level can lead to an overcomplicated solution that you then need to clean up, so it is worth matching the level to the task.
• Dependency. Delegating everything is tempting, so finding a balance matters: review the output carefully and keep your own coding skills sharp.
• Side effects. While debugging, Claude may clear caches, recompile code, or reindex, which can break the local environment, and test data may be left behind. Keep an eye on the commands it runs and check that everything is cleaned up.
• Access rules. A good practice is to review the changed files and commit and push them yourself, since the responsibility for the code stays with the developer. Reading data freely is usually fine, but changes and deletions should need your confirmation. Secrets such as .env, app/etc/env.php or auth.json are better kept off-limits entirely, since everything Claude reads goes into the model's context, and the same care applies to database dumps with customer data. Such rules can be set up as allow and deny rules in the permission settings, which can also be shared with the team through the repository.

Where we stand

Claude Code clearly speeds up understanding a codebase, but we still need time to learn where it really helps and where reviewing its output costs more than it saves. We are now building shared practices for working with AI, so that it fits well into both individual work and the team's workflow. There are also features we have not fully explored yet, such as custom agents, MCP servers, and permission rules shared across the team.

Our advice if you want to try it: keep CLAUDE.md up to date, write clear prompts, and check the result.