The New Worst It Will Ever Be
I've been disillusioned.
When I wrote my last post in July, I had hit a wall with AI coding. It was magical and frustrating at the same time. Most of my journal entries had focused on just how awful the experience can be when you hit context limitations. To try and make it work consistently over large projects, I found myself building tools: flow controllers, state machines, validation hooks. The energy I spent wrestling with the tools became more draining than being productive. At the same time, the thought of coding by hand felt like the stone age.
The Quagmire
Along the way, I consumed my regular feeds of design, tech (here, here, and here), and product creators. I felt relieved watching others hit the same walls. It was validating. Feeling seen, I decided to stop arguing with my LLM and reflect on why the current workflow was so frustrating.
What I missed about the old way of programming was the control. If something didn't work, there was only one person to blame. With AI, even when I did everything right, the output could go wildly off the tracks. The same prompt could give me brilliance or garbage. It was hard to trust.
But the speed at which I could work, wow. That's what made it so hard to walk away. AI could get me to a working outcome in a fraction of the time. As a busy parent working all day, I didn't have hours to burn on side projects. The potential to build almost anything was right there, just out of reach. I could feel it, but I couldn't control it. The probabilistic nature of AI defeated my ability to keep my apps organized. So I had a few options: slow down and code "the old fashioned way," keep hacking through the growing pains, fully embrace the vibes and accept the chaos, or just wait it out.
I chose to wait, watch, and learn.
Sonnet 4.5 Reignites the Spark
Soon after it arrived, I built a tool to help me organize my finances, and it actually worked. I got to the outcome I wanted and replaced old spreadsheets. It felt great to build something to completion and not have it fall apart. I now have a permanent, usable app that I can keep customizing and making my own. Goodbye Google Sheets.
It really felt like a breakthrough. Both Claude Code and Sonnet 4.5 were a huge step forward. The drift bombs weren't gone, but they seemed less debilitating.
Then Came Opus 4.5
In the short time I've been using it, the experience has been nothing short of amazing. It breezes through tasks and evaluates the codebase thoroughly before starting. I've learned to have it reflect back what it's finding: describe what it sees, play back what it hears from me, ask questions before jumping in. But honestly, I haven't needed the elaborate CLAUDE.md files or careful rules anymore. It just seems to know what to do and creates elegant solutions.
I've said "wow" out loud numerous times as it worked through ideas I couldn't crack before. I originally built this site using Sonnet 3.7, and I could never get past generic solutions, especially around animations and nice UI. Opus 4.5 eats it for breakfast. I found myself art directing more and thinking less about prompt perfection. The output felt more authentically me, rather than generic web UI.
Here's the thing: I use VSCode, and I found that I wasn't looking at the code as often, or at all. Claude Code used to be a third of my window and the code two-thirds. Now it was the opposite. I'd open my browser, review the output, give more prompts, and Opus crushed task after task. This felt like a true abstraction away from code. A person with an idea, some historical knowledge of programming, and a grasp of system design could talk to an AI and bounce between prompts and output. Every so often I'd review the code to see if it made sense, and it did. Trust was being built. I felt in control, and "coding" became less important.
Honest Caveats
I should be clear about what I'm claiming here.
I was blocked. Now I'm unblocked. But I've been here before. The honeymoon periods in my earlier entries ended with drift bombs and context collapse. So the question isn't whether this is good. It is. The question is: where's the next wall?
I don't know yet. Maybe it's still the context window, just at a larger scale. Maybe it's something I haven't encountered: maintainability over months, debugging code I never read, handing a project to someone else. Maybe the wall doesn't show up until the project gets genuinely complex.
What I do know is this works for me, today, at this scale. I'm a solo builder making personal tools. I haven't proven this holds up over longer time horizons or with collaborators involved. But it's further than I've gotten before. And that's exciting.
This is the new bar for "the worst it will ever be." In a few months, I'm hoping I'll be writing about the things I've built rather than ranting about the tools again.
Where This is All Going
A new abstraction. A new interface. Code becomes an intermediate language that only AI cares about.
Who looks at intermediate languages anymore? Do you review the JavaScript that TypeScript compiles to? I understand that code is deterministic in a way that AI output isn't, but if the AI can produce things within an acceptable range, it's likely good enough, and perhaps better.
There is a massive value proposition in making software development accessible to more people. More ideas turned into working prototypes means more opportunities for people who don't have the wiring for sifting through thousands of lines of code hunting for a bug.
I imagine the next interface looks something like a system map. You design the architecture, the relationships, the data models, and the AI handles everything below that. It's like visual programming or node-based tools, but one layer higher. "Vibe coding" was all intent and no architecture. This would be intent plus architecture, with the implementation and deployment mostly automated.
You still need to think. You still need to understand systems. But you're thinking about shapes and relationships, not syntax and implementation details.
Wrapping Up
The next twelve months are going to be fascinating. Gemini 3 just landed. OpenAI has declared "code red." Who knows what the next version of Claude will unlock.
For the first time in months, I'm reinvigorated about building ideas rather than fighting with tools. We're in this awkward in-between where the new thing hasn't fully formed yet, and the current state is a Frankenstein of old methods mixed with new ones.
I do agree with my developer influencers that learning to code is still important. Not for the syntax, but for how programming makes you think about systems and the interconnectedness of a lot of moving parts.
If this is the worst it will ever be, I can't wait to see what's next.