What happens when you ask an AI to design your database schema from scratch and just... take its answer?
That question is worth sitting with, because it's happening constantly. Developers paste a rough product description into Claude or ChatGPT, ask for an architecture recommendation, and implement whatever comes back. Sometimes that goes fine. Often it creates technical debt that's expensive to unwind later, because the AI made assumptions the developer didn't notice.
Where AI Coding Tools Actually Shine
Claude and similar coding assistants are genuinely good at generating implementation code. Give them a clear function signature and describe the expected behavior, and they'll produce working code faster than most developers can type. They know common patterns, they've seen thousands of similar problems in their training data, and they don't make syntax errors on the basics.
They're also good at explaining existing code, writing tests against specified behavior, and refactoring small functions. These are mostly local tasks - the context you need is right there in the file or the module.
What They Don't Know About Your System
Architecture decisions require context that an AI doesn't have from your prompt, no matter how detailed you make it:
- How your team deploys and what your infrastructure actually looks like
- The technical debt already in your system and what's too risky to touch
- Your team's strengths - a schema that requires deep PostgreSQL expertise is a bad fit for a team that runs mostly MySQL
- How your product is likely to change in six months based on conversations happening in your company
- The parts of your codebase that are politically off-limits for unrelated reasons
When Claude recommends a microservices split or a particular database schema, it's pattern-matching against architectures that appeared in its training data. It doesn't know your constraints. It doesn't know what you said no to six months ago. It doesn't know that your DevOps team can't support a Kubernetes migration right now.
The result is recommendations that look reasonable on the surface - they'd work fine for a greenfield project at a startup with a full infrastructure team - and might be exactly wrong for your actual situation.
The Right Frame for AI in Software Development
AI coding tools work well as implementation assistants, not as decision-makers. The distinction:
Implementation: You've decided to use a PostgreSQL table with this schema. Have Claude write the migration script and the query layer.
Architecture: Should we use PostgreSQL or a document store? How should we partition this data? Where should service boundaries go?
The second category requires you. Not because AI can't generate an answer - it absolutely will, confidently and in detail - but because evaluating that answer requires context only you have.
Treating AI output as a draft to react to, rather than a solution to implement, is the discipline that separates developers who use these tools well from developers who accumulate AI-generated technical debt. Ask Claude to generate three approaches to a problem. Then decide which one fits your actual situation.
The failure mode isn't AI giving bad architecture advice. It's developers not recognizing they're receiving architecture advice at all, because it arrived in a code block rather than a design document.