In short: ChatGPT helps me every day, but rarely the way people imagine. I don't have it write finished code for a client project and deploy it. AI helps me most before the code: sorting requirements, classifying errors, explaining unfamiliar code and acting as a second opinion. Responsibility for architecture, security and production changes stays with me.
The real work starts before the code
A large part of development is asking questions: what does the client really want? Which existing interface can be used? Which change affects other parts of the system? This is exactly where ChatGPT helps. I have it sort requirements, compare alternatives and phrase checkpoints.
With Shopware extensions in particular, it pays to separate data model, storefront, Administration and version compatibility early on. If I describe the problem properly, with Shopware version, PHP version, desired behaviour and known constraints, I get useful ideas. Without that, I get code that only fits at first glance.
Where AI really helps me
- Classifying errors: going through log messages and symptoms systematically, collecting possible causes and setting a sensible order for checks.
- Understanding unfamiliar code: having existing PHP or Symfony code explained and possible side effects worked out. That saves a lot of time, especially in old, grown projects.
- Collecting test cases: which cases do I need to consider for a change? Empty values, missing permissions, multiple languages, unusual product variants.
- Thinking through SQL and migrations: playing queries and data changes through before they hit real data.
- Translating client requests: turning an informal email into an actionable task list.
- Second opinion: when several technical approaches are possible, I have the pros and cons laid out and check whether I missed something.
Where it gets dangerous
A language model doesn't know which files are on a server, which plugins modify a shop or which customisations a predecessor made. It occasionally invents methods, mixes up Shopware versions or underestimates the consequences of small changes.
The tricky part: the answers almost always sound convincing. But a well-phrased suggestion is not a reliable solution. Typical traps:
- methods or events that don't exist in the Shopware version in use
- solutions that modify the Shopware core instead of using a plugin
- security gaps, such as unchecked input or missing permission checks
- database queries that work with test data but collapse with tens of thousands of articles
Then there is data protection. For client projects, I check beforehand which information may be sent to a cloud service at all. Credentials, customer data and confidential code don't belong in a chat.
A workflow that has proven itself
- Describe the problem: as concretely as possible, with versions, starting point and what must not be changed.
- Ask for causes and approaches: not a big patch right away, but hypotheses and possible solutions first.
- Decide myself: which option fits the actual application, the project and the budget?
- Implement: in small steps, traceable and under version control.
- Check: code review, tests, backup and ideally a test environment before anything goes live.
Why automated tests matter so much here is covered in Unit Testing in PHP. And why security checks must never be skipped in Avoiding Security Vulnerabilities in PHP Code.
And when AI does write code?
For my own projects, I let AI do more. How I built a complete website with Antigravity is described in Antigravity in Practice. The difference: there I carry the risk myself and can check everything at my own pace. Client projects get a stricter standard.
Conclusion
AI makes me faster mainly at analysis, research and preparation. It is an excellent sparring partner, but no substitute for experience with the specific system. Responsibility for architecture, security and production changes stays with the developer. That combination is where I see the greatest benefit.
Frequently asked questions
Can ChatGPT write a Shopware plugin?
It can generate a skeleton and individual building blocks. Whether the plugin fits the version in use, is secure and works together with other extensions has to be checked and tested by a developer.
Does AI really save time in development?
Yes, especially for debugging, getting into unfamiliar code and preparatory tasks. If you take AI code over unchecked, you quickly lose the time saved on debugging.
May I paste client code into ChatGPT?
That depends on your contract with the client and the service's settings. Clarify it beforehand. Credentials and personal data never belong in a chat.