A solo developer building the 'Kotomka' SaaS planner using Claude Code encountered issues with UI consistency and data loss. The project stack includes Next.js 16, Postgres (Neon), and Vitest. To maintain UI standards, the developer implemented a custom audit script that tracks CSS values across 36 screens and 3089 elements, replacing hardcoded styles with design tokens.
The developer addressed data synchronization issues by implementing a version-based conflict resolution system. A race condition where unsaved changes were overwritten during server pulls was fixed by adding a state check before resetting the 'dirty' flag. Synchronization was optimized with a lightweight version check every 2 seconds, replacing the original 15-second polling interval.
To manage AI behavior, the developer reduced the prompt context from 951 lines to a 117-line core, using lazy-loading rules for specific components. Automated guards, such as a PreToolUse hook that blocks unauthorized file creation in the UI directory, were introduced to enforce architectural constraints that the AI might otherwise ignore.
The developer emphasizes that while AI accelerates coding, architectural oversight remains a human responsibility. Automated tests and audit guards are essential, as AI-generated documentation and code analysis often contain inaccuracies. The project is currently participating in the A:START accelerator program.
