Completing an AI-generated project using multiple models , and learning the hard way what AI can’t do.
The Starting Point
I was handed two things:
- A half-built project using TypeScript, JavaScript, and Supabase
- A presentation outlining what the system was supposed to become
At first, it sounded straightforward.
Then I realized something important.
The existing codebase wasn’t written by a developer……. it was generated using AI.
My task wasn’t just to finish a project……
It was to turn a partially AI-generated system into something that actually worked.
The Approach
To complete the project, I used multiple AI models:
- Claude Haiku 4.5
- Codex
- Gemini Flash
Each model helped in different ways:
- Generating missing features
- Fixing bugs
- Speeding up repetitive development tasks
But the real work wasn’t just generating code.
It was making sure everything aligned with the original feature requirements from the presentation.
Where AI Helped
AI made certain parts of development significantly faster:
- Turning feature descriptions into working code
- Handling repetitive logic
- Suggesting quick fixes for errors
At times, it felt like having multiple assistants working in parallel.
The Real Challenge
The challenge wasn’t building features.
It was making everything consistent and cohesive.
Because I was using multiple AI models, I ended up with:
- Different coding styles across files
- Inconsistent logic patterns
- A codebase that worked ,but didn’t feel unified
On top of that, the original AI-generated structure lacked a clear architecture.
So even when features worked, the system didn’t always make sense as a whole.
What I Learned
This experience made one thing very clear:
AI can generate features. It cannot understand your system.
That responsibility stays with the developer.
I had to:
- Interpret the presentation correctly
- Decide how features should connect
- Ensure the system behaved logically from end to end
Without that, the project would have remained a collection of working pieces ,not a real system.
What I Had to Do
To bring everything together, I had to step in as the system designer:
- Review every piece of generated code before using it
- Refactor inconsistent logic
- Standardize patterns across the project
- Simplify overly complex solutions
Only after doing this did the project start to feel stable and understandable.
The Common Misconception
It’s easy to assume:
With enough AI tools, the system will build itself……..
In reality:
AI builds parts. Developers build systems.
The more tools you use, the more important your role becomes not less.
The Right Way to Use AI
This experience changed how I approach AI in development:
- Use AI to accelerate execution, not replace thinking
- Don’t let AI define your architecture
- Always validate output against real requirements
- Prioritize understanding over speed
Final Thought
This project wasn’t just about finishing code.
It was about connecting three things:
- A predefined vision (the presentation)
- AI-generated outputs
- Human decision-making
And turning them into something coherent.
AI can help you move faster.
But only you can make the system make sense.
A Question Worth Asking
Before calling any project done, ask:
- Does this system reflect the original vision or just what the AI generated ………….
Because those two are not always the same.
