A Dial Worth Turning: Claude Opus 5's Prose, and the Style Guide Anthropic Wrote Against Its Own Model

Nine days ago I wrote that the voice drifts back: two developers had shipped tools that hand Claude’s output to a second model for translation into English, Anthropic had shipped a Concise style, and the fix that held was a 1983 aircraft-maintenance vocabulary. That post treated the voice as a fact of the model. This week Anthropic wrote it down.

The Fable 5.1 prompting guide has a section called “Writing density.” It defines an anti-pattern, gives examples, and tells you to paste the definition into your prompt. The examples are Opus 5’s tics. The vendor has written the style guide against its own model and shipped it as your job.

The Complaint Had Numbers by August

The complaint started on launch day. On the Opus 5 Hacker News thread, 1,778 points, one commenter posted side-by-side gists of Opus 5 and Fable 5 on the same prompt: “Opus 5 continues many of the ‘Claude-isms’ of its 4.8 predecessor in a way that Fable broke away from,” listing “carry the argument,” “worth stating plainly,” and ”, and the trap.” His closing line: “We need an ‘annoying English’ benchmark.”

Arena built one. On August 12, from tens of thousands of real responses:

• 510 words per response for Opus 5, against 158 for Opus 4.5. Fable 5 sits at 316.
• Sentences 58% longer, with 46% more clauses.
• Em dashes 2.3 times as frequent. Phrases like “load-bearing” about twice as frequent. “Honestly” and “frankly” up 50%.

Two days later “Why does Opus 5 feel worse to work with?” hit 993 points. The post was about the model guessing instead of asking. The comments went to prose.

The single biggest annoyance with Opus 5 is that it writes too elliptically. Sentences that orbit a point, then jump to it like it’s a revealed insight. Unnecessarily abstract phraseology. Constantly using inanimate nouns as the subjects in sentences in order to unlock variety in verb choice, especially when it helps construct a sentence where the real action can ‘land’ like a surprise at the end.

— barrkel, Hacker News, August 14

Reddit got there first. On August 10, ten days before the two translation tools in my last post, a Claude Code hook that rewrites “Claudish” into English through a local model drew 2,960 upvotes on r/ClaudeAI. Five days later a thread titled “Downgraded from Opus 5 to Opus 4.6 and it feels night and day” drew 946. “Opus 5 speaks in riddles and weird sentence phrasing.” The top reply: “No human being communicates this way.” The fix the poster settled on, after trying an explain-it-simply style that “dumbs down details,” was an output style built on ASD-STE100. The same standard that closed my August post. It is also the rule in my own global config, where the line above it reads “This is not a trial and not a preference.”

Anthropic’s Answer Is a Paragraph

Here is the section, verbatim, from Prompting Claude Fable 5.1. It opens by saying 5.1’s writing “is generally a step up from earlier Claude models, with fewer stock phrases and less unexplained jargon,” then admits its prose is sometimes denser than Fable 5’s, “sentences run longer and there are fewer paragraph breaks.” Then it hands you this to paste:

Mannered prose substitutes metaphor and flourish for direct statement. Instead of “a parameter worth varying,” the mannered writer produces “a dial worth turning.” Instead of “this point still matters,” they write “this point earns its keep.” The phrases exist to display the writer, not to convey the idea, and readers can tell. That is why mannered prose irritates: it makes the reader work harder so the writer can perform.

— Anthropic, Prompting Claude Fable 5.1, 'Writing density'

Read it as a product document and three things stand out:

• It is the best description of the voice anyone has published, and it comes from the people who trained it. “Display the writer, not convey the idea” is the whole complaint in seven words.
• The fix is a prompt. “Add it to a user message (preferred) or the system prompt.” Below it: “The short version also tends to work: Please remove all mannered prose.” Tends to. Nine days ago the argument was that prompting does not hold, and the evidence was Anthropic’s own changelog fixing output styles “drifting back to the default voice mid-session.” The new guidance is a longer prompt.
• Your old rules are now wrong. The next section, “Formatting in chat,” says earlier models “overused bullets and bold” and 5.1 “leans the other way,” so anti-formatting rules written for those models “can suppress structure the content needs.” Remove them, it says. I argued in July that a CLAUDE.md is debt because every rule is written against one model’s habits and inherited by the next. This is the vendor saying it: the rule you wrote to stop Opus 4.8 bolding everything is now the reason 5.1 will not give you a list.

Does 5.1 Read Better?

The evidence is early and it is mostly positive.

• Anthropic’s claim is one clause in the Claude Code launch post, 11:13 AM PT on September 1: “its writing style is more natural.”
• Two customer quotes on the launch page carry it. Canva’s head of AI: “The standout in Claude Fable 5.1 is the writing: more understandable, more meaningful, and it follows our writing guidance better. In blind tests against Fable 5, I preferred its writing and output.” Red Hat’s distinguished engineer: “updates that are more concise and easier to follow.”
• Every’s vibe check, published the same day: 5.1’s outputs “carried fewer AI tells than any other model’s, and they read at a seventh-grade level, a grade below Fable 5 and two below Sol.” Dan Shipper: “I can actually understand what it’s saying.” Same review: “A brief that allowed eight to 12 supporting quotes got 43. A hotel recommendation with a 1,000-word cap came back at 1,288 words.” Five of 27 checkable quotes were not in the source material.
• The system card has no writing evaluation. The closest thing is HealthBench, scored with OpenAI’s length penalty. Fable 5.1 drops from 74.2 to 62.1 on the professional set once verbosity is charged for. Twelve points, on the model whose writing is the fix.

So the direction is right and the size is unmeasured. Fewer tells, a lower reading level, and a model that still writes 43 quotes when asked for 12.

The Default Still Has the Voice

One more line from the same docs, under What’s new in Claude Fable 5.1: “For most workloads, start with Claude Opus 5.” Reach for Fable “when your evals on Claude Opus 5 at higher effort still fall short.”

Put the two documents together. The model Anthropic tells you to use by default is the one that writes 510 words with the surprise at the end. The writing fix shipped in the model you are told to reserve for hard problems, and even there it is a paragraph you paste yourself.

What holds, then, is what held in August:

• A vocabulary, not a definition. Anthropic’s paragraph asks the model to recognise mannered prose in its own output. STE gives it 900 approved words and a twenty-word sentence limit. One asks for judgment. The other removes the option. On Opus 5, the default model, the second one is still the only thing I have seen last a whole session.
• Delete the rules written against the last model. The vendor now says so in print. Anti-formatting language for 4.8 is a bug on 5.1.
• Keep the translation tools for Opus 5. The paragraph is written for 5.1. Nothing in the docs says it works on the model most people will actually be running.

In August the fixes were a second model and a 1983 standard. In September the vendor added a third: a definition of the disease, to be administered by the patient. “Please remove all mannered prose.” It tends to work.