My Evolving AI Coding Workflow: Using Gemini + VS Code Agents to Build Faster

Hey IH fam,

Wanted to share my current workflow for spinning up new projects, leaning heavily on AI. Like many of you, I'm always looking for ways to build faster and focus more on the product itself, especially when bootstrapping or working solo. This process is definitely a work-in-progress, but it's showing some promise for getting MVPs off the ground quicker.

Here’s the current flow:

1. Idea -> Gemini Canvas: I start by outlining the app or system I want to build directly in Google Gemini Canvas. I try to be pretty detailed with the requirements in plain English.

2. AI Generates Design Doc: I then prompt Gemini to take that description and structure it into a system design document.

3. Component Breakdown: Next, Gemini breaks down that design into the necessary components, classes, file structure, etc. – basically creating the project blueprint.

4. Refine in Google Docs: This blueprint goes into a Google Doc. This step is crucial because I can easily iterate, ask Gemini for changes, and refine the prompts before generating code. It saves a lot of back-and-forth later.

5. Initial Code Generation: Once the plan looks good, I have Gemini generate the initial codebase based on the refined blueprint in the Doc.

6. Code -> VS Code: Export the generated project files and open them up in VS Code.

7. AI Agents Build: This is where I bring in AI agents directly within VS Code (using models like GPT-4o or Claude 3.7) in "agent mode," they can autonomously work on implementing features based on the initial plan and code structure. They handle a surprising amount of the initial build-out.

Next Steps & Thoughts:

I'm planning to experiment with incorporating MCPs (Model Context Prompting) soon to see if that streamlines the process even further.

My main takeaway so far is how this shifts my role during the initial development phase. It's less about painstakingly typing every line of boilerplate and setup code, and much more about:

• Clearly defining the requirements upfront.
• Directing the AI effectively through prompts.
• Critically reviewing the AI-generated code for logic flaws and quality.
• Testing the output thoroughly.
• Editing and guiding the AI when it goes off track.

Obviously, this doesn't replace deep architectural thinking or complex problem-solving, but it does significantly accelerate getting the basic structure and features in place. It feels like moving from being just the coder to more of an architect/editor/tester, especially in the early stages.

This flow seems pretty effective for getting the skeleton of an app built quickly.

Anyone else using a similar AI-assisted workflow for development? What tools, prompts, or techniques are you finding effective (or ineffective!) for speeding things up? Curious to hear your experiences!