I have seen many posts recently about the frustration with agent.
A very common complaint seemed to be the following:
- Agent is not great at building complex apps
- When Agent is adding new features, it breaks the old features.
Then you find yourself stuck in an endless loop while slowly dying inside…
If this describes you, this post is for you!
As of early 2025, AI excels at creating simple functions but struggles with building complex applications in one go. Through trial and error, I’ve developed a workflow that significantly improves success rates by breaking down features and building them incrementally.
Planning Phase: Defining MVP
Start by listing all desired features, then categorize them into:
- Serial features (those that build upon each other)
- Parallel features (those that work independently)
Create a tree diagram to visualize feature dependencies. For example:
App
├── Authentication (Parallel)
│ ├── Basic Login
│ ├── Password Reset
│ └── OAuth
└── Data Processing (Parallel)
├── Input Validation (Serial)
├── Data Analysis (Serial)
└── Export Results (Serial)
Building Phase
Key Rule: Don’t include all features in your initial prompt. Instead:
- Create separate Replit projects for each parallel feature
- Build and test each feature independently until functional
- Commit working code to GitHub immediately (trust me, this saves headaches!)
For parallel features with serial components:
- Build the core feature first
- Git commit when functional
- Create a new branch for each additional serial feature
- Test and commit after each successful feature addition
- Merge branches as appropriate
Finally, create a new Replit project to integrate everything. Pull all the working code as context and ask the AI to help combine the features.
Remember: Commit early, commit often. Your future self will thank you!
Hope this helps!
Bonus: if you find this helpful, you might find my recent long form post on AI interesting too.