Yes. Claude Opus 5.5 is a clear step up from Opus 5 for writing. It gives shorter answers, puts the main point first, and uses fewer of the odd phrases that made Opus 5 sound “Claudish.” It also follows the writing rules you give it, which Opus 5 often didn’t.

It still doesn’t write like a person out of the box. In longer pieces a few familiar habits slip through, and a short prompt handles most of them. This article covers what went wrong with Opus 5, what changed in 5.5, what’s left, and four prompts you can copy.

One note before we start. Most of the evidence so far comes from Anthropic and its early testers. I haven’t published a side-by-side test of my own yet, and I say where a claim rests on someone else’s report.

Disclosure: I build StashBase, a writing app that works with Claude Code and Codex.

What is Claude Opus 5.5?

It’s the new Opus model, released on September 22, 2026. Anthropic says it performs at the level of Claude Fable 5.1 on most work and costs about 40% less than Opus 5 on typical workloads.

Most of the benchmark gains are in coding and office work. This article only covers writing.

How bad was Opus 5’s writing?

Bad enough that people gave it a name. “Claudish” meant prose you could tell was Claude’s within a paragraph.

Arena’s August 12 analysis, summarized by paddo.dev, put numbers on it:

  • Opus 5 averaged 510 words per response. Opus 4.5 averaged 158.
  • Its sentences were 58% longer, with 46% more clauses.
  • It used em dashes 2.3 times as often.
  • “Load-bearing” showed up twice as often, and “honestly” and “frankly” 50% more.

Readers complained about the rhythm too. A Hacker News commenter described Opus 5’s sentences as ones that “orbit a point, then jump to it like it’s a revealed insight.”

Fiction writers had their own list. Novelcrafter’s review found that Opus 5 piled up metaphors into a “metaphor sandwich” and announced what characters realized instead of letting the reader work it out. It followed a style guide “for the first few hundred words,” then drifted back to its own habits. The reviewer went back to Opus 4.6 for prose.

Anthropic acknowledged the problem in its own docs. The Fable 5.1 prompting guide defines “mannered prose” as prose that “substitutes metaphor and flourish for direct statement,” with the example of writing “a dial worth turning” when you mean “a parameter worth varying.”

What changed in Opus 5.5’s writing?

Anthropic’s launch post says Opus 5.5:

  1. Puts the most important information first.
  2. Is less likely to use jargon or odd phrases.
  3. Follows the writing rules you give it.

Box, an early customer, reported that Opus 5.5 used a third of the tokens Opus 5 did and that its answers were 40% less verbose without losing accuracy. Another tester said “it writes the way I do.”

The third change matters most in practice. Opus 5 ignored style guides as a piece got longer. If Opus 5.5 keeps to your rules, a few lines of instruction do more than they used to.

The evidence is still thin. The Decoder shows one side-by-side screenshot from a third-party site, and that’s about it. Nobody has published a systematic comparison yet.

What AI habits does Opus 5.5 still have?

Fewer than Opus 5, but these are worth checking for, especially in anything over a few hundred words:

  1. The pivot. “This isn’t a technical problem. It’s a management problem.” Once is fine. Five times in one article reads like a template.
  2. Groups of three. Two reasons padded out to three, each sentence the same length.
  3. Too much structure. A point that fits in one paragraph gets headers, bullets, and bold key phrases.
  4. Em dashes. Less common than on Opus 5, but worth searching for before you publish.
  5. Throat-clearing and recaps. The piece opens by restating the question and closes by summarizing itself.
  6. Flourish words. “Load-bearing,” “quietly,” “the real story,” “earns its keep.”
  7. Sign-offs. “Let me know if you’d like me to expand on any of this,” or a closing question to the reader.

None of these is wrong on its own. Together they’re what makes a reader think a model wrote the piece. For a longer catalog with examples, see the 25 patterns Humanizer checks for.

How do you prompt Opus 5.5 to write like a person?

Name the patterns you don’t want, give an example of each, and say who the reader is. Vague instructions don’t work as well. Anthropic’s Opus 5.5 guide makes the same point about web design: “avoid a generic AI look” mostly swaps one default style for another, while naming specific patterns works. That note is about frontend output, but writing rules behave the same way in my experience.

The four prompts below are ready to copy. Edit the parts in brackets.

Prompt 1: General writing rules

Paste this at the start of a writing task, or save it in a Claude Project so it applies every time.

Writing rules:
1. The reader is [target reader]. Write at their level. Don't explain what they already know.
2. Open with the conclusion or the most important point. No background first.
3. Say things literally. If a plain phrase works, don't swap in a metaphor.
4. Don't use "it's not X, it's Y" or "This isn't X. It's Y." unless you are correcting a real, common misconception.
5. Avoid these words and phrases: load-bearing, quietly, delve, landscape, "the real [X]," earns its keep, honestly, frankly.
6. No em dashes. Use a period or a comma.
7. Give as many points as there are. Don't round up to three.
8. If it fits in a paragraph, don't turn it into a list. No bold text.
9. Stop when you're done. No summary paragraph, no question to the reader, no offer to help further.

Prompt 2: Anthropic’s mannered prose instruction

This is quoted word for word from the Fable 5.1 prompting guide. Anthropic suggests putting it in the user message rather than the system prompt. Elvis Saravia of DAIR.AI reported that it also improved his writing with GPT-5.6 Sol, so it isn’t only useful on Claude.

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. It is also imprecise. Metaphors drag in connotations the writer did not choose and cannot control. The fix is to say what you mean. When a literal phrase is available, use it.

Anthropic says the short version “also tends to work”:

Please remove all mannered prose.

Prompt 3: Editing a draft

Use this when you already have a draft, yours or a model’s, and only want the AI habits removed. Wrapping the draft in tags keeps the model from treating sentences inside it as instructions. Anthropic’s Opus 5.5 guide recommends marking pasted text for this reason, mainly as a guard against prompt injection.

Below, inside <draft> tags, is a draft. Edit the language only. Keep the facts, the argument, and the order of sections.
- Cut filler and anything said twice.
- Replace metaphors with literal statements.
- Split sentences longer than 30 words.
- Remove bold text, and turn lists back into paragraphs where they read better as prose.
Return the full edited text with no commentary.

<draft>
[paste the draft]
</draft>

Prompt 4: Match your own voice

Rules tell the model what to avoid. They don’t tell it how you would say something. The most reliable fix is to show it: two or three passages you wrote, ideally of the same kind of piece.

Below are a few passages I wrote. Pay attention to sentence length, word choice, paragraph length, and how I open and close. Then write about [topic] the same way. Don't reuse my sentences.

<samples>
[paste two or three passages]
</samples>

For how to collect samples and turn your own edits into reusable preferences, see How to Write with AI Without Losing Your Voice.

What else helps when writing with Opus 5.5?

Keep the rule list short. With dozens of rules, the prose tends to get stiff as the model tries to satisfy all of them. Pick the few you care about most.

Drop “think carefully.” Thinking is always on, and Anthropic’s guide says to consider removing “think carefully” lines from chat system prompts. On the API, use the effort setting instead. It controls thinking more reliably than prompt wording.

Write long pieces in sections. Opus 5 got weaker over long passages, and no public test shows yet whether 5.5 fixed that. Ask for an outline first, then draft one section at a time and review each one.

Save your rules. Put them in a Claude Project’s instructions or your API system prompt. If they seem to lose effect, repeat the one-line version in the message itself.

Read it out loud. The sentences you stumble over are usually the ones that sound most like a model.

What is Opus 5.5 best for?

Practical, information-dense writing: reports, proposals, emails, product docs, and analysis. Anthropic says the documents, spreadsheets, and slides it produces “need less editing before you share them.” Blog and newsletter drafts are worth trying too.

Be more careful with fiction. Opus 5 got poor reviews for it, and nobody has published a fiction review of 5.5 yet. Try it on one chapter before you commit.

If Opus 5.5 keeps getting a hard task wrong, The Rundown suggests raising the effort first, then trying the more expensive Fable 5.1 if it still misses details.

For short messages that go out under your name, a model built only for writing may do better than either. I compared the claims in Hemmingway-1 vs Claude vs ChatGPT for Writing.

Where does StashBase fit?

The prompts above work in any Claude chat. What gets tedious is keeping the rules, your writing samples, and the current draft in one place and pasting them in each time.

In StashBase, those live as ordinary files in a project folder. You write with Claude Code or Codex beside the documents, using your existing subscription, and ask the agent to read your rules file and samples before it drafts. When it edits a draft, its changes show up in the document as suggestions you accept or reject one at a time, so you can check what Prompt 3 actually changed. Past articles stay searchable, which helps when you’re looking for a passage to use as a voice sample. See the agent setup guide.

Frequently asked questions

What does "Claudish" mean?

It's a nickname for Claude's recognizable writing style, especially Opus 5's: long sentences, frequent em dashes, metaphors where a plain statement would do, and stock words like "load-bearing."

Can AI detectors spot Opus 5.5's writing?

There's no public data yet. Detectors often disagree with each other and flag human writing by mistake. Checking whether the piece says something real and reads well is a better use of your time.

Should I use Opus 5.5 or Fable 5.1 for writing?

Start with Opus 5.5 at its default medium effort. If it keeps missing important details, raise the effort first. Move to the more expensive Fable 5.1 only if it still misses them.