Hello! This is part 1 in (hopefully) a series where I, an AI doomer, am exploring how AI can be practically applied to my software work.
I’ve historically been firmly in the anti-AI camp for software engineering. My assessment was that AI models are essentially just confident bullshit-artists - and I’d like to think I value correctness in my work highly - so I mentally classified them as tech-bro-hype and didn’t really give much more thought to it.
My workplace is pretty conservative as far as AI adoption goes too. We don’t have a clear policy on approved AI usage, and nobody is advocating very strongly for AI adoption within the business. A few colleagues have messed around with using GitHub Copilot for cracking out boilerplate-y test code but that’s about it.
However, I stumbled across this post last night just before bed and it made me seriously stop & think. Honestly I barely slept because my brain wouldn’t shut off.
First - the author’s experience of seeing how people interact with AI matches mine closely. I think the chatbot-interface that LLMs present causes people to not take them very seriously; learning how to extract value from an LLM isn’t treated as something worth studying.
Second - the idea of controlling LLM outputs via rules like this & building up a set of project-specific rules over time is just brilliant. Without having tried it yet, I feel like this is a total paradigm shift - particularly when it’s combined with the approach where your AI agent can run tests & build your code to validate the correctness of its work.
Third - I would very much like to be employed in 5 years time, and exploring what this tech is capable of feels like a no-brainer for self-preservation at this stage.
So I’ve come up with a plan of attack which goes like this:
- Install Cursor and switch from PyCharm (sadge) to it for side-project work.
- Starting with MPC Autofill, have a crack at using this rule-driven AI agent workflow for developing in an established project. I’m thinking I’ll tackle the following in roughly this order:
- Overhaul my project’s wiki.
- Identify areas lacking in test coverage and add tests.
- Identify areas where code cleanliness could be improved and refactor them.
- Have a crack at implementing a new feature - a system for managing in-flight and completed image downloads.
- Be critical of my workflow and the quality of the model’s outputs.
- Check in with another blog post at this stage & review the experience.
- If all is going well - take this to the next level with a more complex side project & look at integrating Cursor into my professional work.
So here we go! Feeling pretty excited to be honest.