55 entities. 78 API endpoints. 114 UI components. Over 100,000 lines of code. And it took about 20 hours of AI generation time.
A development team recently documented how they built an Aircraft MRO (Maintenance, Repair, and Overhaul) management system - a serious enterprise application - almost entirely with AI tools. The result is one of the more honest assessments of where AI coding actually works and where the popular "vibe coding" approach falls apart.
Vibe Coding Hits a Wall Around 10K Lines
Vibe coding - the practice of prompting an AI to generate code conversationally, fixing errors as they come up, and iterating until things work - is genuinely useful for prototypes, small tools, and personal projects. But this team found it fundamentally cannot scale to enterprise complexity.
The reason is architectural. When an application has 55 interconnected database entities and 78 API endpoints, no developer can hold the full mental model in their head while prompting an AI line by line. Vibe coding is bottom-up: you build pieces and hope they fit together. Enterprise software demands top-down thinking: you design the whole system first, then build the pieces to specification.
Anyone who has tried to vibe-code past a few thousand lines has felt this. The AI starts contradicting its earlier decisions. Data models drift. You spend more time fixing integration bugs than you saved on initial generation.
The Blueprint Method
Instead of prompting their way through the application, the team used Gemini to create a detailed architectural blueprint with a human-in-the-loop review process. This blueprint defined the data model, API contracts, UI component hierarchy, and business logic rules before a single line of application code was generated.
Once the blueprint was locked, they handed it to Claude Code, which generated the full codebase in roughly 20 hours of autonomous execution. The key insight: Claude Code was not doing creative problem-solving. It was executing a well-defined specification. That is a fundamentally different task, and one that current AI tools handle much better.
This two-phase approach - AI for architecture planning, AI for code generation, humans for review at each gate - produced a working enterprise system at a speed that would have been unthinkable two years ago.
What This Means for Your Projects
The lesson is not that AI coding tools are inadequate. It is that they need structure to work at scale. For small projects under a few thousand lines, vibe coding remains the fastest path from idea to working software. For anything with real complexity - multiple data models, role-based access, business logic with edge cases - you need a blueprint phase.
The practical takeaway: treat AI code generation like you would treat a junior developer. Great at executing clear specifications. Unreliable when asked to make architectural decisions across a large codebase. The teams that figure out this division of labor first will ship faster than everyone else.