AI-Driven Development When You Kind-of Sort-of Know What You Are Doing
There was a time in the early 2000's when I had servers in my basement and understood the LAMP stack well enough to prototype applications as a single developer working alone. This was back in the days of Myspace and Geocities, when normal people learned HTML so they could express themselves with unicorn wallpaper and flying toasters on personal websites.
I was reflecting the other day that sometime toward the end of that decade I stopped prototyping things online and shifted to other activities. The reasons were many, but could be summed up in a few basic concepts:
- the technology we were using, as it got more capable, got a lot more complicated, to the point where it was difficult for someone whose day job was not full time coding to maintain any sort of competence as a full stack developer.
- The rise in difficulty meant that even professional developers started to strongly specialize into frontend, backend, database, etc...
- which meant that even smallish projects started to require teams, and to get going those teams required money...
- which mostly came from venture capital and private equity, who are famously not very tolerant of the frivolous...
and just like that what had once been a fun world full of personal artistic expression and 'all your base are belong to us' became a place where every single thing you created or interacted with was somehow optimized to result in monetizable human behavior, and absolutely nothing more....

When we started the web the key challenges were bandwidth and hosting. The technology was simple enough to figure out even without a lot of online documentation, but you needed a server somewhere to host your stuff, and that server needed enough bandwidth to actually content to browsers, which cost real money.
The challenge now, in whatever Web N version we are now on, is complexity, and the use of complexity as a moat to prevent the siphoning of attention from existing powerful actors. Have an idea for a simple web site that interacts with people in a meaningful and creative way? To make that happen you'll need to get really good at things like React, and Tailwind, and Node, and Firebase, and Git and a long list of other things that are not really optional and not fully comprehensible to the moderately technical person who sort-of kind-of knows what they are doing. For someone who does not do it a lot and doesn't having existing templates and code they are comfortable with just lying around, taking a Sunday afternoon just to set up a project like that seems pretty exhausting to contemplate.

Enter AI-driven software development. I'll admit for a long time I have been skeptical of this. The skepticism was founded in the thought paradigm I was exposed to: a developer writing lines of code gets suggestions from an AI in the IDE on how to make those individual lines of code better. In this paradigm the developer is still expected to be an expert in the language and components the code is being written in and the technologies being interacted with, and is expected to exercise judgement on whether to accept the suggestions the AI proposes. Since the AI gets it wrong a significant percentage of the time, the developer becomes a defensive actor preventing the spoilage of the code base due to inexpert AI contributions.
This did not seem helpful in addressing the challenge of how to make the idea of a Sunday afternoon web site development session seem like a lot of fun. So I ignored it.
The paradigm being presented here is completely different. Set up Cursor as an AI IDE. Set up a hotkey so you can talk - yes vocally with your mouth talk - to the IDE. Turn on what is called YOLO mode where the AI can make any necessary changes to any file on your computer. Focus on the result, not the code, or even necessarily the technology selections. Press that hotkey and start talking.
This is a completely different paradigm. Here you are only required to be an expert in describing the outcome you want, and, like a well trained soldier, giving just enough instruction to get what you want. Let the AI figure it out. If it gets it wrong, make it do it again, until it gets it right...
I know many senior developers (I work with very competent senior Engineers every day) who complain that this approach would only dilute the quality of their code base and its performance in a very undermining ways. This is probably true. The bigger complaint is that this paradigm essentially turns them into product managers, as a hands-off-keyboard approach essentially reframes their contribution to one of oversight, not execution. Also, probably true.
But here is the thing: for someone who kind-of sort-of knows what they are doing and wants to spend Sunday morning prototyping things at their kitchen table, all of that is extremely liberating. Most of the value a person like that brings to the kitchen table is a vision about realizing some imaginary thing and a strong detail-orientation about how that thing should work. There is no value gained by internalizing the details of the technology at this early stage. The value is in quickly realizing the vision, and testing it for appeal.
As a test of this approach, I installed Cursor and took it for a spin. I used it to build from scratch a new version of my personal website at http://www.malcolmstanley.com. To make it fun, first I built a fun version of the website, themed like a personal page from the late 90's:

Then I asked the AI for a second view of this content that is clean and minimal and modern:

When I was done, I saved it to Github, and deployed it, basically for free, to Vercel. You can see it right now.
What I found was really encouraging, with caveats. There were some things I wanted to do that required multiple attempts by the AI coder to achieve. As predicted by the senior engineers, the AI screwed up the code quite often, and then required multiple attempts each time to get it working properly. But I did not care, because I got to issue instructions like 'The PURPLE, THE PURPLE, MAKE IT GO AWAYYYY" and then watch the AI do in a few attempts over a few minutes what would have taken me hours to even develop a strategy for how to do it.
(when I typed that last line I tried to make the text that says PURPLE actually be purple but you cannot do that in the default Ghost blogging interface. This is what I meant when I said 'somehow optimized to result in monetizable human behavior, and absolutely nothing more...' Ghost is great, but its no Myspace...)
More than anything, the whole PURPLE thing felt like a lot of fun, and I realized I had really missed that about working with web tech. I had forgotten how great it could feel to make something simple and fun and release it to the world.
So now I have a list of prototypes I want to try and make, things that last week I was unaware were even frustrating or annoying or exciting that I wanted to fix or try.
So my early conclusion is that YES, AI is dangerous as hell and could end humanity, but also its not going away anytime soon, and like anything that dangerous, with the right mindset, maybe we can have a lot of fun with it while we are waiting for it to destroy us. I've turned on YOLO mode, and I'm not turning it off.
Cue the fun.
About Me
I'm a full lifecycle innovation leader with experience in SaaS, ML, Cloud, and more, in both B2B and B2B2C contexts. As you are implementing your value proposition, I can take a data driven approach to helping you get it right the first time. If that seems helpful to you, please reach out.