AI & Robotics Lab

💻 Using AI for Coding - Part 2: Workflow

Previous part: Choose the White or Black Box

In this chapter I focus on how to effectively use AI as your "programmer buddy." The main idea is that while AI knows tremendously much, only humans can set meaningful goals and choose the right path to achieve them.

When working on a project, my workflow follows these steps.

1. Architecture Consideration
Start by discussing ideas and the overall project architecture without writing any code. The outcome of this stage is a Block Diagram of the main logic. For all subsequent stages, it's very useful to keep this diagram in the model's context.

2. Defining Project Structure
Still without coding, discuss how to best divide functionality between modules and classes. This produces a structure of main modules and their relationships - either as a diagram or text description. This structure should also remain in the model's context during later stages.

3. Main Logic Implementation
The first step involving actual coding implements only the basic skeleton:
• Main logic with class drafts
• Error propagation
• Logging
• Testing framework
• Metrics handling

With AI assistance, I select the best patterns for code organization. This is also the time for research to explore new technologies and approaches. The result is a working prototype that demonstrates both the core functionality and crucial supporting elements like error handling and logging. The first basic tests also appear at this stage.

4. Main Development
Only after the prototype proves it can properly solve the task do I begin the main coding phase. Here I finalize the project structure beyond the main modules and classes, and create a proper test suite with high coverage.

In my experience, tests are best developed after a module or class is nearly complete. This requires a good playground to run and debug all elements separately - which is why creating a proper structure with modularity, logging, and error handling in the previous step is so important.

AI is very helpful at this stage for answering specific questions and quickly finding the causes of errors. While it's tempting to let AI create tests, there's a catch: AI can produce professional-looking test suites, but you may spend much more time making them work properly than if you created them yourself one by one, only consulting AI for specific questions.

5. Deployment
It's important to think about deployment from the early stages: will the project be deployed as a binary, a package, or perhaps simply cloned from GitHub without building? This consideration helps organize code more effectively.

This overview describes how I'm trying to use AI effectively for developing code in a "white box" manner. AI is tremendously helpful, but it can't eliminate fundamental design stages: architecture planning, block diagrams, prototyping, and release iterations remain essential across all engineering disciplines - mechanical, electrical, software, and beyond.

What do you think about this approach? I'd be pleased to hear your thoughts in the comments!

⚡2

38 views08:48