Lalit Maganti, a Senior Software Engineer at Google and maintainer of the Perfetto tracing tool, wanted to build better developer tools for SQLite for eight years. He finally shipped syntaqlite after three months of evenings-and-weekends work. The difference? Using AI coding tools like Claude Code on the Max plan at £200/month. Maganti tried what he calls the "maximalist" approach first: delegating almost all design and implementation to Claude. The result was functional but fragile. When he reviewed the codebase in late January, he found "complete spaghetti" with functions scattered across random files and some files grown to several thousand lines. He threw it all away and started fresh, this time taking ownership of architectural decisions and using AI as "autocomplete on steroids" rather than a semi-technical manager. The core technical challenge was parsing SQL exactly like SQLite does. SQLite has no formal specification and doesn't expose a stable parser API. The only solution was extracting parts of SQLite's C source code, which contains over 400 grammar rules in a dense style. This is what killed the project for years. It was hard and tedious, with no guarantee of success. Maganti credits AI with helping him overcome procrastination. Code generation accelerated dramatically. Claude even generated over 500 tests he could reuse. But architecture emerges from how components interact, and that requires human oversight. If you're using AI agents, they're great at local code generation. You still need to own the big picture.