Should you be “vibe coding?” And if so, what does that mean, exactly?
This question isn’t for the early adopter tech bros who are already building the next OpenAI using OpenAI to make a better OpenAI. This question, taken seriously, is for me and you, the normies.
Me? I’ve already got decades of software development experience, starting from just before the internet became the internet. I helped build and ship one of the first generative AI platforms that wrote billions of automated, informative articles in 2010 and is still working today.
You? You might be even more skilled than me, not gonna lie. You might be an entrepreneur, a CEO or a rock star innovator at whatever it is you do. Or you might just be hiding in a cubicle somewhere, trying to figure out if the opportunity cost of wrapping your brain around these latest AI advancements is just going to end up getting you laid off anyway.
I got all your backs.
Look, I still hate calling it “vibe coding.” But lately, as AI code assist tools have gotten better — and by “better,” I mean that they’ve gotten more restrictive in their approach to coding — I find myself using them for more and more coding tasks that I wouldn’t normally have.
So yes, regardless of how much tech is in you or your business, from a very little to a whole lot, it’s time to hire your own personal AI tech lead and put them to work immediately — if you’re not already going down that road.
With some limits. You’re gonna want to set limits.
Why an AI Personal Tech Lead Makes Sense
I’ve been using AI coding helpers out in the wild on various gigs, for sure. Honestly, though, I’ve been using them mostly to reduce the amount of time I spend with my hands on my laptop keyboard and to increase the hell out of my productivity when I do pound away at my laptop.
You should be using AI in the same way that you should already be stringing together pre-existing tech to help you do your job, whether you’re a full-time coder building a full-fledged digital product or CEO of a service-oriented company building some light automation tools around your customer support flows. Or anything in between.
If you haven’t already taken the plunge, you should definitely be diving in cautiously. But no matter how much AI is in your diet these days, I’ll document some of the biggest drawbacks that still frustrate me about these AI code assistant tools.
If you already consider yourself a vibe-code hero, these same tips will give you some things to think about in terms of what AI makes it easy to get lazy about, until things break.
First and foremost, AI can do a lot of things for you. It can build entire apps, product lines, even an entire business operation from scratch. What it won’t do is spin value out of nothing. You still need to be the driver of the value train, the keeper of the proprietary data, processes or other intellectual property that customers will pay to access.
So, you should be coming at this like AI is your personal tech lead. And by “personal,” I mean it should be helping you do those techie, keyboard-initiated things you would do on your own if you had the time and the inclination to learn the skills.
For example, a lot of my original vibe coding went into getting myself out of the death spiral.
How AI Can Help You Slow the Death Spiral
The death spiral is when you spend so much time executing the unrelenting work that needs to be done to keep the business operating at full speed that you don’t speed enough time — or any time — stabilizing and growing the business. Then one day, the work inevitably becomes more than you and your infrastructure can handle, and the business comes crashing down.
I used AI coding helpers to get me out of that spiral. Don’t forget that your laptop is a small but powerful server, with the ability to automate a lot of those high-effort, low-reward repetitive tasks that are clogging your growth.
And that’s where a personal tech lead can help.
To be crystal clear here, I’m not talking about creating an all-knowing, 20-handed AI agent from scratch. I’m not even talking about “building” generative AI or tapping into the dangerous power of LLMs. I’m just talking about using AI and its code assist tools, like Claude Code or Copilot, to help you automate custom workflows to do your dirty work. I have a post-publish process that, as my writing career grows, ballooned to 90 to 120 minutes, three times a week. After applying automation, I got that down to five minutes.
Simple. Or look, if you want to build the next Shopify, you can do that too.
Either way, here’s what to watch out for on that path to AI Vibe Code Utopia.
If You’re Coding in Public, You Need a Babysitter
And this is the same advice I would give myself, even with my old-school skills.
I know enough to know when code or a solution strategy is either stupid or broken, and I’m just doing a lot of work for myself and some MVP stuff for clients. But even then, I get nervous about the security, privacy and scalability holes that creep in under my nose.
Until you feel like you know you’ve got everything the code does nailed, you should stay local. Like I said, your laptop is a server, and I’m running back end servers from PHP to Python plus a database and even a few third-party helper apps locally. This means the code never leaves my laptop, and thus, I don’t have to worry about security or privacy beyond my own machine.
When I do code for the public internet, much as I have awesome editors, I also know a lot of developers I can reach out to. And now that I have a couple “real-code” apps out in public (I’d been no-coding with Bubble the last few years), I’ve got two of my developer friends on retainer, just to babysit my Git as I smash the keyboard like a monkey on a mission.
A Vibe Coding Method: Discuss -> Build -> Review -> Ask
Here’s a simple working method for getting started with vibe coding.
A 4-Step Process for Vibe Coding
- Discuss.
- Build.
- Review.
- Ask.
Discuss
Before I do anything, I want to have a discussion with Claude as to how it can be done. I tell it what I want to do, and ask it to discuss how we might do it. This will also help me narrow down my prompt into something that won’t do things I don’t want done. Claude and his AI counterparts aim to please, and will suggest, attempt or even just do all sorts of things you don’t ask for in the name of making your life easier.
Build
Whatever you’re building, whether it’s a helper, a feature or an entire app, you’ll want to start with the smallest and simplest of steps, but something more than “Hello World.” If you have AI build the very first thing a user (you) sees and takes action on, that’s probably more than enough. Get the small steps right so you don’t get overloaded or miss big gaping holes in logic.
Review
You don’t have to understand each bit of syntax (although you should), but you’ll want to know enough to know what you can and should rewrite yourself and change. At some point, you want to tell Claude to start spitting out changes, not rewrite entire files. And you’ll want to be able to implement those changes yourself.
Ask
Ask for help getting your back end and front end working. Oh, also have Claude comment thoroughly. Do not forget to ask for this. Ask it about what it did and why, over and over again until you understand it. Again, don’t skip this step. I can’t tell you the number of times I’ve “borrowed” real code from somewhere only to realize it was 100 percent wrong for what I needed.
See, with both no-code and AI assisted code, you’re not forced to know what you’re doing until you are, then by that time you’ve built so much around a step you did long ago that turns out to be junk that neither you nor Claude will be able to fix it and maintain your progress.
Finally, if you’re not 100 percent crystal clear on anything I’ve written here, rather than me writing 5,000 words to address multiple different scenarios, you can ask your personal AI tech lead your specific questions as they related to what you want to accomplish.
The Technical Debt Struggle Is Real
Way back when I first started using the AI code helpers, I got into technical debt in five minutes. Now, more than a year later, it’s still hard to avoid inelegant solutions and spaghetti code.
AI will “try” a lot of things. It will fail often. Then it will keep apologizing and correcting itself with confidence and still be wrong. In a lot of ways, it’s like working with some of the not-so-great developers from my past. It’s not that they don’t hear me. It’s that they push stuff out based on what they understand because I didn’t explain it well.
Ask AI to clean up after itself often, but lead that process with what you know and have learned along the way, so it doesn’t clean you out of a process you needed or open a security hole.
The best ounce of prevention I've found to reduce technical debt and inelegant solutions is the “Discuss” part of the method above. Use the discussion to let it give you options you might not be not aware of, like a third-party extension that might do exactly what you want. Ask while you discuss, but don't execute until you understand.
Be Super Specific About What You Ask For
And that’s not prompt engineering. It’s about understanding and breaking down what you need and when before you ask for all of it at once.
Claude is a yes man. If you don't tell your AI what NOT to do, you’re risking it doing something it thinks you want. Again, not unlike a human developer. The more time and energy and discussion and asking you do before you get to the execution, the better your execution is going to be. That’s … easy to say.
I haven’t played around with agents as much as I'd like, but having AI remember the discussion and those questions and being able to refer to specifics from them can save you a lot of time and headache. It also reduces the need to write a super specific prompt.
That said, work your way to more long term projects. You should NOT be spending $200 a month like you would with Claude Code until you’re absolutely stopped out and you can see the ROI on your investment.
In fact, I’ve used AI a couple times to get to a point where I’ve decided that what I’m working on shouldn’t be automated in the first place, or at least not in the way I’m thinking.
All this leads to a quick final discussion about “magic.”
Limits Are Everywhere, and They’re Good
Claude and his counterparts, by nature, have a lot of short-term limits on what you can accomplish at each step along the way, both in terms of how much it can do for you and how much you still have to do by hand. But this is by design, as it can tell you to manually do much more than you would want to let it do itself. That said, especially in the last six months, a lot of those limits can be overcome. As agentic AI becomes more mainstream, you’ll be able to code your way through a lot of current AI code assist limitations.
Also, remember, at any time, the providers can artificially limit what you can do.
You’ll eventually have to get into the code, even if just because, at some point, someone will need to dive into hundreds or thousands of lines of code to make small changes or tweaks. Nothing is free, not even AI-based automation. Be wise, and remember, if you automate something and it breaks, you should still be able to and remember how to do it manually.
In the end, the wrong way to approach AI coding is to dream up some magic and tell AI to make it happen. What you should be learning is how to break down “magic” into smaller and smaller elements that you can understand and build your own magic around.
Claude gave me a lot of help organizing and building tools to help with my business. It did not, and will never, be able to automate the “magic” that makes what my business offers valuable to my customers.
