Why Ai Works For Me

Why AI Works For Me

I read a reddit post today asking Is Claude Writing Better Code Than Most Of Us?.

For me, the answer reminds me of Matt Lucas’ “Vicky Pollard” character in Little Britian, whose catchphrase is “Yeah, but no, but yeah, but no, but…”. Let me explain.

TL;DR

For enterprise coding there is a floor to the level of quality necessary, LLM’s exceed that because it has enough training data to do it, and is far faster than most human devs.

Longer Story

Background

For me, and I answered in the post, the answer is “Per hour, absolutely.” My tool of choice right now is a terminal/TUI with Claude Code so using the Anthropic models, although this has nothing necessarily to do with that particular tool.

My world is “enterprise” development. That is, writing primarily web based systems with well known languages (Javascript/Typescript) with a separate backend also with well known languates (Java, Ruby, Python, bash, SQL) using well known deployments (AWS, Kubernetes, Github Actions) with well known monitoring (too many to list). You see a pattern here.

VERY little in my world is “novel”. It may be a step above CRUD apps, but not much. We use established patterns and idioms to glue together well known frameworks in expected ways to get some bespoke output based on the data we have.

Value

Our value is in the Product Managers knowing what our particular customers want to do, and we write code to allow them to do that, hopefully with as little friction as possible. People like me generate the most value by delighting the customers as quickly as possible.

This is an important thing though, and there’s more to it than just shipping code quickly. Some of that is saying “no”. It’s not a delight to customers if your design is slowing you down. It’s not a delight if you have to support and work around code for features that no one uses. It’s not a delight if your build process takes hours. So it’s a balance of getting out the right stuff and not letting the wrong stuff get in your way. This is mostly design and not code/implementation, which is where devs like me who have been doing this for too long have some extra value.

Right, And?

So how does AI factor in here? Because almost all of my “work” is using well established tech, all the LLM vendors had an enormous corpus of prior art to steal, errr, “train on”. So the models well know all the patterns, frameworks, best practices, error conditions, etc. It’s generally VERY GOOD at coming up with a reliable, idiomatic solution for your question. This all does require some good prompting, context window management, etc., but we’re all learning that. And, it does this in minutes where it would have taken me hours or days of (very education/experience) trialling and erroring and reiterating. And, it’s generally good enough. It’s readable, idiomatic, has acceptable error checking, etc. It’s not exciting code, but the older I get the less I want that.

Would my artisinal, hand-tooled, anguished over code be “better”? Probably, for some value of “better”, but for this particular type of application, that doesn’t matter and LLMs are a hell of a lot faster. The “quality / minute” is just higher. Design does still matter, a lot, and that’s where I and my colleagues still do it mostly by hand (or by whiteboard).