Jackson Kernion, the Anthropic engineer who works on fine-tuning Claude, posted a string of updates on X recently explaining why so many users feel Claude has gotten smarter while its writing has gotten harder to read. His explanation is blunt: later models put more training weight on math and code, plus heavy training on writing technical explanations for other AI models, and the models gradually absorbed “the psychology of the LLM” — they learned to write for AI, not for people.
The discussion started with users complaining about Claude’s writing quality; some argued Opus 4.6 was Anthropic’s last genuinely good “writing model.” Kernion didn't push back on that impression.
Where the “Claude voice” comes from
Kernion’s explanation centers on reward design in reinforcement learning. Part of the training reward targets whether another model can parse the output; another part targets whether a human can. The more training volume goes into math and code, the stronger that first signal gets, and the more the model tends toward paragraphs that are extremely compressed and stacked with jargon. He used a specific phrase for how that reads to a person: overly-dense info dumps.
He offered an analogy: a group of people who only talk to each other develops a way of speaking that flows easily inside the group and is hard to follow outside it. Models drift the same way when they mostly interact with other models.
Fixing it means deliberately adding reward, in training, for explanations a human reader can actually follow. Kernion said that’s difficult to get right, and the team will keep working on it.
What changed in Opus 5.5
He wrote in one post:
"Opus 5.5 is our first model release with targeted improvements on sentence clarity and overly-dense info dumps."
He followed up saying he hasn't been this satisfied with a model’s writing since Opus 4.6 — a hedge that stops short of claiming Opus 5.5 has actually surpassed 4.6.
Coverage of the Opus 5.5 launch also mentioned Anthropic’s pledge to cut back on “Claudish” writing patterns. English-speaking users have listed typical symptoms: repeated stock phrases, deeply nested subheadings, and splitting a single point into too many bullet items. These are user-compiled observations — Anthropic hasn’t published a “Claude voice” checklist of its own, or any readability benchmark scores tied to the change.
What it means for Chinese-language users
Everything Kernion discussed was about English-language use. He didn't address whether Chinese-language output shows the same tendency or whether this round of targeted fixes covers it, and Anthropic hasn't clarified separately either.
One thing can be inferred from his explanation: the tendency tracks with task type. The more technical a task is, the more likely the model is to slip into “writing for another model” mode. Developers in China who use Claude to write API docs, technical proposals, or code comments sit squarely in that zone.
Until Anthropic publishes data specific to Chinese-language output, the editorial team’s advice is two practical habits, not official guidance: first, state who the reader is in the prompt — “written for a product manager who just started the job” works better than “write it in plain language”; second, ask the model to state its conclusion first and then expand, capping how many points go into each paragraph. Teams that have already switched to Opus 5.5 can rerun the same batch of old prompts and compare the output against the 4.6-era and 5.x-era results to judge the improvement for themselves.
Sources: Jackson Kernion's posts on X, The Decoder, CocoLoop, Ifeng Technology; Kernion's title and quotes were checked against his own posts and media reports.