The Real Vibe Coding is Finally Here

Making coding accessible to anyone was the biggest focus for AI models this year. Now, it's finally starting to feel real with Opus 4.5, Anthropic's latest release.

If you haven't heard of vibe coding by now, just take a minute to scroll through Twitter - stories of people building fully working prototypes in minutes or products over weekends are easy to find. Using platforms like Replit, v0, Lovable base44 or bolt, people can describe what they want and get a functioning piece of software working with AI.

But actually try it yourself and you start hitting a wall: A bug that won't get fixed (despite the AI claiming it's gone). Asking for simple changes that never quite hit the mark. Hidden dependencies and things you never asked the AI to do that are now causing problems.

The issues compound until you start playing a roulette game of 'try another prompt' to see if maybe, this time, it'll finally work. Hundreds of dollars of API calls later, you're navigating a codebase you don't understand, trying to undo whatever is keeping you stuck.

I've scrapped entire repos once I realized I was never going to get clarity on what was happening. The AI was allowed to run amok, creating slight drifts here, a small deviation here, until what I got doesn't match what I thought I had asked for. At one point I realized Cursor was completely replacing files instead of updating them, despite my prompt saying to 'update'.

It wasn't all bad. I built some stuff. A simple prototype showing how we can use Git-like commands to collaborate on knowledge work at scale. A semi-working proof of concept to turn text notes into Gantt charts and timelines. A barebones product landing page with a working email signup.

But honestly, I found these embarrassing. I had a deeper need for data manipulation that none of these concepts and prototypes could touch. The user interface and interactivity weren't the problem - the problem was I needed a customized data pipeline, manipulatable features to feed models, metrics to measure outcomes, vector and graph based databases that pulled across everything to build the vision that I had for a product I've been incubating since last year.

Essentially, I needed the ability to build an actual AI product, not another traditional SaaS that would have quickly outlived its purpose once you rethink its workflow with AI.

So I didn't have much hope when I setup VS Studio Code and Claude Code a few weeks ago, driven by an urgent need to solve a personal problem with automation. It was less complex than my product vision, and I knew the technical features were well within feasibility.

I had also learned by now (painfully), that AI vibe coding tools were not necessarily good planners. I gave myself the time - this time - to really work out exactly what it is that I needed to build, with the help of AI chatbots that were good at brainstorming and fact checking.

Once the requirements were set, I just had to feed it to the agent. This was my first time with Claude Code, but after hearing engineers repeatedly recommending it I got over being burnt out by tools that under delivered and gave in. The launch of Opus 4.5 was a coincided blessing as I was able to experience Claude Code with the latest model.

I'm now convinced that THIS is truly the moment that vibe coding actually works. Especially for audiences like me, who know what they want and can speak to the technical level, but not parse exact functions and code-speak.

What's been surprising isn't the fact that it was quick to build a working prototype - that's been figured out. It's the fact that I am continuing to build and iterate on it over weeks without encountering major issues.

It's not perfect - the model has to guess when to make assumptions and when to ask for clarity, and that still causes drift from intention. But the same thing happens with human coworkers.

If only this was how I get to work all the time.