AI-boosted,
Not AI-led
A few words on the most hotly debated topic of AI, and how we use it in our processes.
AI-assisted Development
While not universally useful or perfect in any way, agentic AI coding tools have matured significantly from where they were in 2023-2024. We are now firmly in the realm of AI writing reliably functional code 99% of the time in absolutely inhuman quantities.
The downside, as always, is that the more you use it without paying attention, the more your code devolves into a gigantic avalanche of slop that, despite being written in a human-readable language, becomes so deranged, that it effectively ends up being unreadable for any sane person who wants to dive in and make manual changes.
But... unlike just a few years ago. This code is actually perfectly functional once you start testing all the features you just vibed. Which is certainly not nothing.
What that means is that there is an expanding category of tasks for which manual intervention might mot be necessary at all.
So if we were to sum up our AI usage in software development, we would split it into 3 very distinct categories:
Mostly Vibes
Mostly Vibes
Our Blog Page Builder that we ship with some of our projects is a great example of this approach. At first, we were trying to follow the code and refactor the slop as it was piling on. But at some point we figured, 'Why are we doing it?' The page builder has no security implications, it's in its own corner of the overall app, which is fairly isolated from everything else. As long as it works, we're good. And it just does.
Also, the thing about slop, is that even though barely readable by humans, it is something that AI itself can actually make perfect sense of. So, as long as we are not running into model's context window limits (which are very wide with Gemini), we're also good. Because we can continue to ask an agent to add more features on top without having to ingest all of the underlying code madness ourselves.
AI-assisted
AI-assisted
This is for parts of our sites that have some business data or security implications. While agents can handle things like auth and CRUD (database) operations almost flawlessly these days (provided clear prompts and instructions); we absolutely cannot let it just rip and have to closely follow what's happening in the code. A single bug in those critical parts could make authentication unsafe and/or lead ot unpredictable behaviors or changes to the database when performing simple actions. Which could be a disaster waiting to happen later when the site is already live.
So, with this middle-of-the-road approach, we do let AI do a significant portion of writing the code, but because of constant human assessment of the progress, it can only go as fast as the speed of human comprehension.
Largely Human
Largely Human
While AI today is near-perfect for smaller-scale backend tasks, it's still completely all over the place when it comes to building visuals.
AI and vibe-coding bros will try to convince you that vibin' an entire website in seconds is the future, but the reality is much more uneven. And literally every time we tried to completely replace the entire visual development side with pure vibes, it backfired spectacularly in the long run.
As of right now, creating precise, responsive, and meaningful layouts is still a very demanding task, requiring a lot of human attention to do right. And more often than not, we prefer the manual control over it.
However, since we can't completely ignore AI in that area either, we decided to embrace the all-or-nothing approach and created a special Turbo package that mostly relies on vibe-coded pages. It's a nice way to 'something' up and running very quickly.
AI-assisted Assets
The great thing about AI-assisted workflows is that depending in the situation, project, and the timelines at hand, you can actually be very flexible with how much manual effort you put in.
Media Manager
There are many instances and projects where “some content” needs to be present (similar to where stock photos could be used in the past). Think - you and your team in certain clothes, or in a somewhat non-descript location, and things like that. In which case, a direct generation straight from a prompt + AI upscale could be a great solution that would take the least amount of time.
In some situations, however, when the scene is more complex or requires a specific pose and camera angle, pure gen AI night not be enough. It tends to generate very generic shot compositions on its own. But if we supply a 3D base to generate on top of, we can easily overcome those limitations and get a pose/angle that is much closer to what we envisioned. The great thing is - we often can get away with only giving it a schematic representation of the scene, with simplified mannequins and gray boxes. It won't be a 100% match, but it will get the idea across.
And finally, for shots that truly require the ultimate character and scene consistency along with realistic poses and camera, we can just go ahead and provide a full 3D render that has all the key details and generate on top of that.
It certainly takes a lot more time and effort. And it probably wouldn't make sense to invest so much time in an image on a regular local business project when other methods already produce quite impressive results.
But on a higher-budget, custom scope projects, this level of control could still come in handy or even make or break the message that the website and marketing materials are trying to convey.
Still Deciding?
Check These:
OR