AI as a scribe for writing code

Lately I’ve thought of AI as a scribe, that I dictate short instructions to and it writes the code for me, saving me the manual labor of typing all of the code. The difference between this and normal AI coding is that you are the brain, the AI is the scribe.

My hands and wrists are not in good condition to say the least, so it makes a difference to me if I can type less. The current generation we have of Claude Sonnet 3.7, OpenAI o1 and DeepSeek have clear limitations. I know from extensive use what these limitations are and what questions its likely to give a low quality answer for. And the context length being too short for any codebase, certainly for Sonnet at least.

But what if we know the answer and just tell the AI what to write? We can achieve more fine-grained results that closely match what we wanted, instead of hoping the output is in the correct style you wanted. I find this method of working to be best in the early stages of the code project, as you would expect. There’s all the boilerplate code to get out the way. The codebase is at zero, so there’s no issues with context length. The way I have always liked to write code is by opening an existing project I made and copy pasting boilerplate over to the new one to get me jump started. I don’t enjoy writing boilerplate, that novelty has long worn off.

With AI, I can quick start by saying hey, I want this standard boilerplate, in this simple coding style, using this architecture. For example in C# I might say, use clean architecture, use Entity Framework, .NET 9.0, here’s the Postgres connection string, and so on.
For Rust I’m not very good with Rust so I deliberately use another strategy which I will elaborate on in another article soon, and that is to tell the AI that I am a dumb person who likes simple and easy to understand K.I.S.S. (keep it simple stupid) code. This prevents it from casual use of advanced features and syntax that I’m unfamiliar with.

In terms of time savings, I get a big benefit. I’m not a quick thinker and I think of all outcomes even when I don’t need to. By using AI, I can avoid overthinking. If you don’t know what to write and don’t know what to tell the scribe to write, then you should have a discussion with the AI about it. Open a new seperate conversation and outline your goals and try and solidify what you’re trying to do, without code. You may want to tell the AI “dont answer with any code examples”. Then finish up with asking for a brief summary of the project, and paste that into a new conversation to start coding. Get it to write some code you can run right away, and start iterating.