Claude got worse, and I miss the golden age.

There was a period where using Claude for code felt illegal in the good way. Not because it was perfect. It was absolutely not perfect. It would still hallucinate imports, invent a library function with the confidence of a man wearing boat shoes indoors, and occasionally hand you a solution that looked correct until the runtime punched you in the throat.

But it had taste. That was the thing. It could hold a codebase in its head for long enough to make a decent structural call. It would notice when a problem was actually caused by a bad abstraction two folders away. It could say, in effect, "your request is wrong, fix this instead," and be right often enough that I forgave the arrogance.

That version felt useful in the way good tools are useful. A little sharp. A little dangerous. Occasionally brilliant. You had to supervise it, but you were supervising a capable goblin, not a customer support intern with a React badge.

Now? Now Claude too often feels like it has been laminated.

The product around it is better. I am not blind. The apps are nicer. The workflows are better. The integrations are less embarrassing. Artifacts, projects, desktop things, agentic little rituals, all of that is genuinely useful. Anthropic is clearly building an actual product instead of a text box with vibes taped to it.

Good. Fine. Congratulations to the adults in the room. Have a biscuit.

But for pure coding, especially Opus-style "think hard and help me change the shape of this thing" coding, something has gone limp. It writes more polite code. It explains itself more. It apologizes in more professional typography. It is better at sounding like it has a plan and worse at having the uncomfortable little spark that made the old Claude valuable.

The failure mode changed. That is what bothers me.

Old Claude failed like a brilliant weirdo. It would make a leap, sometimes too far, sometimes exactly far enough. New Claude fails like a committee. It preserves options. It rounds edges. It gives you a reasonable path that slowly calcifies into nonsense after step four. It is less likely to be spectacularly wrong, maybe, but more likely to be quietly mediocre. That is worse for coding than people admit.

Coding is not just typing functions. The typing part was never the magic. The magic was judgment. Knowing which file matters. Knowing when not to patch the symptom. Knowing when the user is asking for the dumb fix because the real fix is harder to phrase. Knowing that a seven-line change can be more valuable than a fresh architecture diagram with all the sincerity of a hotel conference room.

When Claude was good, it helped with that. Not always. But enough.

Now I find myself fighting the model more. It wants to be helpful in the most HR-approved way possible. It wants to summarize. It wants to hedge. It wants to turn a bug into a process. It wants to give me a checklist when what I need is for it to look at the damn stack trace and stop being spiritually employed by Deloitte.

And yes, some of this is probably my tolerance changing. The first time a coding model solves a nasty problem, it feels like witchcraft. The fiftieth time it half-solves one, you notice the smell. Novelty wears off. Standards go up. I get that. I am not pretending nostalgia is a benchmark suite.

But I also do not buy the neat little story that everything is simply getting better in a straight line. It does not feel like that from the chair. It feels like the frontier got product-managed. Safer, smoother, more consistent, more broadly useful, and somehow less alive where it mattered most to me.

Maybe this is the trade. Maybe the weird, brilliant coding assistant had to become a stable platform. Maybe the thing I miss was partly an accident: a model with too much raw edge before the safety rails, routing layers, product incentives, and enterprise-friendly polish sanded it down into something procurement could invite to lunch.

I hate that this is a reasonable explanation. It is very rude of reality to make boring answers plausible.

So yes, I still use Claude. Of course I do. I am annoyed, not stupid. The tools are useful. The app ecosystem is improving. It can still do real work. But I miss the golden age of coding Claude, when it felt less like a platform and more like an accomplice.

I do not want a model that only behaves. I want one that can code with taste, argue with me correctly, and occasionally scare me by understanding the system better than I explained it.

Bring back that Claude. Keep the nice apps. Keep the tools. Keep the grown-up product shell if you must. But put the teeth back in the coding model, because right now the thing is wearing a very expensive cardigan and calling it progress.