Building Syntaqlite: A Journey with AI Coding Agents

For years, I dreamed of creating a set of high-quality development tools for SQLite, a crucial database engine in the tech industry. Despite its importance, no one seemed to invest in enhancing the developer experience for SQLite. This changed when I decided to build Syntaqlite, a project that took about 250 hours over three months, largely thanks to AI coding agents.

In my work with Perfetto, a SQLite-based language for querying performance traces, I noticed the lack of reliable tools like formatters and linters. Existing solutions were either unreliable or inflexible, prompting me to consider building something from scratch. However, the complexity and tedium of parsing SQL like SQLite, which lacks a formal specification, made it a daunting task.

AI coding agents, which I had been using since 2025, became pivotal in overcoming this challenge. Their capabilities improved significantly, allowing me to delegate design and implementation tasks. By using Claude Code, I managed to extract a parser from SQLite sources, create a formatter, and support both SQLite and PerfettoSQL extensions.

However, the initial codebase was a mess—spaghetti code that was fragile and unsustainable. This led me to start over, switching to Rust for better integration and maintainability. I took full ownership of the project, using AI as a powerful tool for rapid prototyping and refactoring, while focusing on design and testing.

AI helped me overcome procrastination by providing concrete problems to solve, rather than abstract design challenges. It excelled at generating standard code quickly, but struggled with non-obvious parts like the parser architecture. The speed of AI also made refactoring easier, allowing me to maintain a clean codebase.

AI served as a valuable teaching assistant, compressing learning time for new concepts and helping me bridge gaps in unfamiliar domains. It enabled me to complete a comprehensive project with features like editor extensions and a WASM playground, which I might have deprioritized otherwise.

Despite its benefits, AI had its downsides. It encouraged a feedback loop of tiredness and procrastination, and I sometimes lost touch with the codebase. AI's lack of temporal understanding and design intuition meant I had to remain engaged and make key decisions myself.

Ultimately, AI was a force multiplier for implementation but not a substitute for design. It accelerated the project but required careful management to avoid pitfalls. This experience taught me the importance of balancing AI's capabilities with human judgment in software development.