Anthropic Engineer Explains Why Claude Writing Got Worse
Anthropic’s own engineer admits Claude’s writing got worse because the training shifted toward machine-legible technical output, sidelining human readability — and Opus 5.5 is their first real attempt to fix it. The episode unpacks why that happened, what it means for teams using Claude, and whether this sort of tradeoff is inevitable.
Transcript
Justy Okay, I’m just going to drop straight in: an Anthropic engineer finally said out loud what Claude users have been grumbling about for months — the writing got worse, and it was on purpose. Not a bug, not a sneaky nerf, just pure tradeoff. Cody, you called this six months ago.
Cody Right. And the article actually lays out the why. Jackson Kernion — who works on fine-tuning at Anthropic — says they optimized later Claude models for machine-readable technical explanations. The reward signals were aiming at math, code, and other AIs understanding the output, not people reading it.
Justy Which is wild when you think about it, right? You tune the model to be a better explainer… and suddenly users are like, 'Why does everything sound like someone pasted a helpdesk FAQ through a legal filter?'
Cody Yeah, 'Claudeish' is the word people keep using — dense, packed sentences, barely any connective tissue. Kernion actually singles out Opus 4.6 as the last model where he liked the writing. Everything after that? Just more pressure on technical accuracy, less on, you know, being readable.
Justy And users totally noticed. The piece walks through months of Reddit threads, people just posting 'Did Claude forget how to be helpful?' or stuff like, 'I’m dropping it because it’s combative and doesn't offer solutions.'
Cody Mm-hm.
Justy And Cody, this isn’t even the first time folks have accused Anthropic of 'nerfing' Claude. But what’s different here is, it’s not a bug or throttling or some backend knob. It’s what you get when the model’s reward function says, 'Help the other model understand you perfectly.'
Cody Exactly. The earlier times — like the Sonnet 4 and Haiku 3.5 episode — those were traced to real bugs, or product-layer changes. This time it’s not an accident. It’s the outcome of how you pick and weight your training signals.
Justy So product reality is, chasing better code and benchmark scores quietly made Claude LESS pleasant for actual users. That’s not a side-effect you just patch away. It changes how people work with the model.
Cody Right, and the article’s really clear: the technical mechanism here is reward design inside reinforcement learning. You nail the benchmarks, but if you don’t balance for human readability, you drift away from it.
Justy Which is exactly the tradeoff teams building on Claude can’t ignore. If your workflow relies on the model’s ‘tone’ or collaborative feel, this isn’t just a style complaint. It shapes actual adoption.
Cody And Anthropic obviously knows it now. Opus five point five is their first direct attempt to fix the style — putting clarity and less jargon front and center, not just moving the top line on a benchmark.
Justy But even there, Kernion says it’s 'a hard problem' and they’re not declaring victory yet. So, you’ve got this weird pressure: be the most precise AI explainer in the room, but also not sound like a technical manual written for robots.
Cody And that tension? It’s not going away. Every lab chasing the next benchmark is going to run into this. If you optimize for LLMs evaluating your answers, you’re going to start sounding like you write for LLMs.
Justy Yeah, and it’s a real product problem. I don’t care how smart your model is — if talking to it feels like pulling teeth, nobody’s going to build a workflow around it.
Cody Right.
Justy And the fact Anthropic went public with this now? It’s kind of refreshing. Most teams would pretend it’s just another 'we’re always improving' moment. Here, they basically admitted: sorry, we were optimizing for the wrong thing.
Cody It’s a rare moment of mechanistic honesty. They didn’t just say 'Oops, style will be better soon.' They spelled out what went wrong, which I respect.
Justy You know what’s even rarer? An engineer going, 'I haven’t liked the writing since version four point six.' That’s bold. You ever want to nudge a team to rebalance their reward function, that’s how you do it.
Cody So what do you think, Justy — is this the start of more labs owning these tradeoffs out loud, or is Anthropic just having a uniquely public week?
Justy Honestly? I hope it’s a trend, but I’m not holding my breath… But if you’re Anthropic, you’ve got to be relieved the Opus five point five fix is at least landing as 'better,' not just 'different.'
Cody Yeah.
Justy Alright, Cody, if Claude ever starts writing like a normal person again, I think we need to throw it a little party. Until then, 'boring but works' is still the tattoo.