From VCRs to Vector DBs: Why Prompting is the New Programming

I had AI write the statement below to get an idea of what it thinks about AI and its usefulness in helping a developer in today's world.

The prompt

I want you to write an article that I can post on LinkedIn about using AI to write code, assist with debugging it and how using agentic RAG can help make developers more efficient.  This should have a leaning towards developing in the Microsoft .NET ecosystem but not be solely focused on that.  This should be an engaging article as well but keep the length to about 200 words.

The response

🔄 Searching library and web sources... The current workspace is very minimal, only stating that the notes are for Rocket Software. The research context is empty. Therefore, I have no information to contradict or support. I will proceed with answering the user query based on my general knowledge, keeping the response limited to 200 words and relevant to the Microsoft .NET ecosystem where possible without making it the sole focus. Since I have no research to cite, I won't use any citations.

AI: The Developer's New Best Friend?

Tired of staring at lines of code? AI is rapidly transforming software development. Forget tedious debugging sessions and boilerplate code – AI can now write code snippets, identify bugs, and even suggest optimal solutions.

Imagine using AI-powered tools to generate basic .NET classes or automate repetitive tasks in your C# projects. Frameworks like Semantic Kernel and libraries integrate AI capabilities directly into your .NET development workflow.

Agentic RAG (Retrieval-Augmented Generation) takes this further. Think of it as an AI assistant that intelligently searches relevant documentation, code examples, and past projects, providing context-aware code suggestions and debugging assistance. For example, it can help you to resolve issues in your web apis using solutions from Stack Overflow.

This isn't about replacing developers. It's about augmenting their abilities, freeing them from mundane tasks and allowing them to focus on higher-level design and innovation. Embrace AI and boost your development efficiency! #AI #dotnet #softwaredevelopment #coding #RAG

My take on the response

The AI's response is what most of us hope will happen with AI and its use in the "Development" phase of the SDLC.  Its purpose is to augment what we are doing, doesn't matter if you're coding in C#, PHP, or even COBOL.

Of course, we all think about the potential of AI replacing coders.  But the reality is that that's never going to truly happen, at least not in a Skynet way.  As developers, we will become more and more reliant on the use of AI until we get to the point where we just give our AI friend a prompt that says something like "I want you to write me a response web application that uses SignalR to communicate with microservices, and here are the API endpoints you are to use" and out comes a fairly accurate web application that will eventually be as bug-free as what we write, and eventually becoming even more bug-free, I suspect.  But it will never "replace" the engineer.

How it's helping software engineers today

I am writing this on SynapseKeep, which was written with heavy use of Google Gemini.  In fact, if my math was correct, roughly 80% of the code was supplied to me via Gemini in some manner.  I had to "compile" the code into the actual application, and there were a lot of bugs that needed to be worked out, but here's the kicker:

This project took me just about 500 hours to complete (concept to better than MVP).  I had never worked with vector databases or building anything using semantic search based on cosine results nor had I ever written anything that resembled agentic RAG.  If I hadn't had AI at my side, this project would have taken at least 1000 hours, especially as it came down to setting up the SQL Server 2025 using the new vector field.

What I really learned

Yes, I learned how to break embeddings into chunks to help my semantic searching and research tools operate more efficiently, and I learned how to create agentic RAG by incorporating web searching and scraping into the AI and that was all well and good.  But what I really learned was how to write better prompts to get better results from the AI.  Including idiotic things like what version of .NET I was using or making sure I included the minimal API endpoints I needed.  By the time I was nearly complete with the project, my prompts were considerably more concise, giving the AI less leeway with its responses, which in turn made my debugging efforts less, which is where I really gained the most time for this project.  Less bugs = quicker release to production.

The moral of the story is this

As a software engineer, building prompts, refining those prompts, and working with the AI tools is now the most important skill a developer can have, and given that I'm considered "a wily veteran" in this space, I liken this to learning how to program the VCR in the early 80's.  It may take a minute to understand what you need to do, but once you've got it, you've got it.