|
I know I was.
I’ve talked about AI a lot this year. If you’re tired of hearing about it, that’s fair. The reality is that I'm personally interested in the topic and it is affecting every aspect of our jobs.
If you're learning to code and want a non-AI article that I think will help you out, check out: https://brianjenney.substack.com/p/if-youre-over-30-and-trying-to-learn
A lot of the advice floating around about AI already feels outdated. Not because the tools changed overnight, but because we keep repeating ideas that never actually worked particularly well in the first place.
Some of these takes were things I personally got wrong. Others are ideas I never really believed, but I still see them passed around constantly.
Either way, they usually lead to the same outcome: frustration.
Here are some of the mistake-iest mistakes you're probably making.
Personified prompts
Why did we do this?
I never believed telling an AI “you are a senior staff engineer” made it smarter. And yet, I still see this advice everywhere in prompting cheat sheets and LinkedIn posts.
These prompts mostly just make you type more.
What actually works is concrete context.
Fix X.
Here is the file where I think the issue lives.
Here are related files showing how we do this elsewhere.
That’s it.
No roleplay. No personas. Just scoped tasks and context.
"AI is only good for greenfield work"
This one I absolutely said.
I was half right.
Greenfield work is still where AI shines. Going from zero to one is insanely fast. I’ve scaffolded entire apps in days that would have taken weeks by hand.
The code was not pretty, but it worked and it looked convincing enough to demo.
Where I was wrong was assuming that legacy codebases and mission critical code were basically off limits.
I recently joined a company with millions of lines of code spread across well over a hundred repositories. I’m learning Kubernetes, Terraform, and more AWS than I ever planned to. Small changes in one service can create big downstream effects.
Vibe coding is not an option here.
AI is not going to magically 10x my output in this environment. But with enough context, it can still help.
Explaining unfamiliar services.
Summarizing logic in languages I don't work with.
Helping with small, well scoped fixes where I already understand the system.
The gains are smaller. The margin for error is thinner. But useful work still gets done.
Not having instruction files for your codebase
This one is completely on me.
On one of my personal projects, I’m doing things that fall well outside the distribution of what most models were trained on. I’m building agents with structured outputs, Python serverless functions on Vercel, the Vercel AI SDK, and newer scraping libraries.
Cool kid shit.
PS. If you're a full stack developer who wants to transition into AI and learn by building, apply for our next 30 day AI Engineer Cohort.
Even though this project is relatively small, under a hundred files, AI stopped being helpful very quickly without guidance.
Claude repeatedly used older versions of libraries I didn’t even have installed. Agents ignored schemas. Structured outputs were handled incorrectly, breaking my chat interface or just not working at all.
At first I thought the tools were bad.
I truly hate to admit this but... it was a skillz issue dawg.
Once I added instruction files explaining the patterns I use, why certain things exist, and what not to touch, output quality improved.
I think of these files as onboarding docs for an developer with amnesia. Because that’s exactly what your AI assistant is.
Using old tools
I'm not one of those AI model bros who uses a dozen different models depending on the task but if you've only tried Cursor using GPT-4, you might want to consider exploring Opus with Claude.
Model quality matters. Tool behavior matters. Some tools are far more aggressive about generating files or rewriting code than others, which can be actively harmful in real codebases.
You’re probably not going to 10x your productivity. But you can absolutely reduce friction, which tends to matter more.
This won't age well
I've re-read many of my posts, articles and watched some of my old videos on AI.
I realized that my opinions would change as the tools and our workflows changed.
I just had no clue how fast this would happen. It's genuinely hard to keep up.
Most frustration with AI does not come from the tools being bad.
It comes from missing context or using outdated advice.
Vague prompts lead to vague output. Missing constraints lead to wrong assumptions. No guidance leads to chaos. Different models can feel like night and day.
If AI feels useless to you right now, you’re in good company.
You might just be wrong about why.
|
