Most prompt-engineering advice this year has been about agents, tool calls, and harness design. This one is refreshingly narrow: a single, short rule about writing style that its author says measurably improved output on two different frontier models. AI educator elvis (Elvis Saravia), founder of DAIR.AI, surfaced it on September 8, 2026, crediting Anthropic for sharing it days earlier — and reported it improved his writing not just with Claude Fable 5.1, but with GPT-5.6 Sol too.
TL;DR
| Question | Answer |
|---|---|
| What does it target? | "Mannered prose" — metaphor and flourish substituting for direct statement |
| Who shared it? | Anthropic, surfaced publicly by elvis (DAIR.AI founder) on September 8, 2026 |
| Which models does it help? | Reported gains on both Claude Fable 5.1 and GPT-5.6 Sol |
| Where do I put it? | As a standing rule in your system prompt or writing/editing instructions |
| Is there a way to test it? | Yes — an interactive comparison tutorial at academy.dair.ai |
| Does it work for all writing? | Best suited to non-fiction, technical, and explanatory writing where clarity beats ornamentation |
The full prompt text
Here is the rule as shared, worth pasting directly into a system prompt, a CLAUDE.md-style instructions file, or a standing custom-instructions field:
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.
That's the whole thing — no elaborate framework, no multi-step chain-of-thought scaffold. It's a single paragraph naming a specific failure mode and giving a specific fix.
Anthropic also documents a five-word shortcut that tends to work when you do not need the full definition in context every turn:
Please remove all mannered prose.
The long paragraph is better when the model is drafting from scratch or editing prose where it needs to recognize the pattern in its own output. The short line is enough for a revision pass on text that already exists — a second-pass edit instruction rather than a standing style constitution.
Why Fable 5.1 made this prompt necessary now
This is not a generic "AI writes badly" complaint. Anthropic's own Prompting Claude Fable 5.1 guide names writing density as a model-specific behavior: Fable 5.1's prose is generally a step up from earlier Claude models, with fewer stock phrases and less unexplained jargon, but in some cases it runs denser than Fable 5 — longer sentences, fewer paragraph breaks, and more figurative phrasing substituting for plain statement.
That combination is why Anthropic published a named anti-pattern rather than telling users to "be more concise." Density and mannered prose are related but not identical. A model can be concise in word count while still writing "a dial worth turning" instead of "a parameter worth varying." The prompt targets the figurative habit specifically, which is the part readers notice as performative rather than merely long-winded.
explainx.ai covered the launch context in Claude Fable 5.1 and Mythos 5.1, where Anthropic framed the release partly as a response to developer complaints about recognizable Claude verbal tics. The mannered-prose instruction is the copy-paste fix for one slice of that feedback — the slice where metaphor replaces precision.
Why this works: naming the failure mode specifically
Most "write better" prompts fail because they're too vague — "be concise," "avoid fluff," "sound natural" — instructions a model can nominally follow while still producing the exact writing style you were trying to avoid. This prompt works differently: it names a precise, recognizable pattern (substituting a dressed-up metaphor for a plain statement), gives two concrete before/after examples, and explains why the pattern is bad (it serves the writer's performance, not the reader's understanding) rather than just asserting that it's bad.
That specificity is exactly the difference explainx.ai's own master prompt-engineering guide emphasizes between prompts that change model behavior and prompts that just add words to the context window: a model can only reliably correct a habit it can concretely recognize in its own draft output. "Mannered prose" gives it that recognizable target.
Why it transfers across models, not just Claude
The genuinely interesting part of elvis's report is that a prompt shared by Anthropic, using Anthropic's own framing, improved output on GPT-5.6 Sol as well — a different lab's model, trained on different data with different fine-tuning. That's evidence the underlying writing habit this prompt corrects isn't a Claude-specific quirk. It's a broadly shared bias across frontier models trained heavily on internet text, where ornate phrasing statistically correlates with "sounds like good writing" in training data, independent of whether it actually communicates the idea more clearly.
This is the same territory explainx.ai covered in Claude Opus 5's load-bearing "Claudisms" — the recognizable verbal tics a specific model tends to produce — except this prompt targets a stylistic bias that appears to generalize across labs rather than being unique to one model's training.
If you are trying to spot AI-generated text by eye, mannered prose overlaps with several patterns in explainx.ai's top 10 signs of AI-generated text — hedge stacking, rule-of-three lists, and the "it's not just X, it's Y" construction are cousins of the same training-data bias toward prose that signals quality rather than delivering it. The prompt is essentially asking the model to self-edit those tells out before you have to.
User message or system prompt — Anthropic's placement rule
Anthropic's documentation draws a practical distinction:
| Placement | When to use it |
|---|---|
| User message (preferred for one-off tasks) | You want this specific reply — a customer email, a changelog entry, a blog section — written without mannered prose |
| System prompt (or CLAUDE.md equivalent) | You want the same style rule on every turn in a long session or across an application that calls the model repeatedly |
The difference is scope, not magic. A user-message instruction affects the current request directly; a system-prompt rule persists across turns so you do not have to paste the paragraph into every conversation. For Claude Code sessions that mix coding and documentation, a standing rule in your project instructions file is usually the right default — the model switches between commit messages, README edits, and inline comments in the same session, and all of those benefit from the same anti-flourish constraint.
A worked before/after edit
The DAIR.AI Academy tutorial walks through a concrete customer-update example. One draft sentence reads:
Automatic imports give your budget a clearer runway, while manual entry remains a path you can continue to walk.
"Runway" and "path" add no information the reader could not get from literal phrasing. After applying the mannered-prose rule, a direct revision might read:
Automatic imports can save you time. You can still enter every expense by hand if you prefer.
The revision states what the feature does and confirms manual entry remains available. The prompt does not require every sentence to be short — it requires each phrase to add information. That distinction matters when you are editing technical docs where some concepts genuinely need a longer explanation; the rule is not "never use metaphor," it is "when a literal phrase is available, use it."
Pair it with Fable 5.1 formatting rules, not old anti-formatting prompts
Fable 5.1 changed more than sentence-level style. Anthropic's prompting guide notes that earlier Claude models overused bullets and bold in chat, which led many teams to add aggressive anti-formatting rules ("no headers, no lists, plain prose only"). Fable 5.1 leans the other way — it uses bold less and is less likely to reach for headers or lists even when the content would benefit from structure.
If your system prompt still carries legacy anti-formatting language from Fable 5 or Opus-era workflows, consider replacing it with a positive rule that says when formatting helps, such as Anthropic's suggested line:
Use lists and bullet points when asked to, or when the content is multifaceted enough that they help with clarity. If the person explicitly requests minimal formatting, always format your responses without bullet points, headers, lists, or bold emphasis, as requested. In conversational, personal, or emotional exchanges, keep to plain prose.
Mannered-prose removal and formatting guidance solve different problems. Running both — direct literal prose and appropriate structure when the content is genuinely multifaceted — produces clearer output than either rule alone, especially for explainers and how-to content where readers expect scannable sections.
When the rule hurts more than it helps
The prompt is optimized for non-fiction, technical, and explanatory writing where clarity beats ornamentation. It is the wrong default for:
- Creative fiction and marketing copy where a controlled metaphor is the product
- Brand voice work where a specific tonal register is the goal, not plain statement
- Poetry, speeches, and narrative essays where figurative language carries meaning the literal phrase cannot
For those tasks, either omit the rule or narrow it: "avoid mannered prose in technical sections; preserve the brand voice in customer-facing copy." The failure mode you are correcting is substitution — metaphor standing in for a literal phrase that would communicate the idea more precisely — not metaphor in general.
Copy-paste templates for common tasks
Customer or product update (full rule in user message):
Write a ~120-word update to [audience] about [change]. Cover what changes, what stays the same, and what action they should take.
Follow this style rule when writing:
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.
Second-pass edit on existing draft (short rule):
Revise the text below. Please remove all mannered prose — replace figurative phrases with literal statements wherever a literal phrase communicates the idea accurately. Do not change factual content.
[paste draft]
Standing rule for CLAUDE.md or project instructions:
When writing or editing prose (README, docs, commit messages, user-facing copy): avoid mannered prose. When a literal phrase is available, use it. Metaphors that do not add precise meaning should be replaced with direct statement.
For deeper prompt-structure context — where standing rules sit relative to task context and retrieved files — see explainx.ai's context engineering vs. prompt engineering guide and the Claude Code token efficiency write-up on what belongs in persistent instructions versus per-turn messages.
How to actually use it
- Add it as a standing rule, not a one-off paste. elvis describes running it everywhere he uses AI for editing or writing — the value compounds when it's always active rather than something you remember to invoke occasionally.
- Pair it with your existing system prompt or CLAUDE.md. It's a style constraint, not a replacement for task instructions — see explainx.ai's guide to context engineering vs. prompt engineering for where a rule like this belongs relative to task-specific context.
- Test the before/after yourself. elvis published an interactive tutorial specifically so people can compare output with and without the rule rather than taking the improvement claim on faith — a good habit for any viral prompt claim, not just this one.
- Expect the biggest gains on explanatory and technical writing. A rule favoring literal statement over metaphor will do less for creative fiction, where a controlled metaphor is often the point, and more for documentation, editing, and non-fiction — the categories where mannered prose most often gets in a reader's way.
- Re-read the output once for density, not just metaphor. Fable 5.1 can still produce long sentences even after mannered phrases are stripped. If readability is the goal, pair this rule with an explicit ask for paragraph breaks on multifaceted explanations — another behavior Anthropic documents separately under "formatting in chat."
How this compares to "write like a human" advice
Generic humanization prompts ("sound natural," "write like a person," "avoid AI slop") give the model a goal without a testable criterion. The model can satisfy "sound natural" while still producing "this point earns its keep" — because ornate phrasing statistically correlates with "good writing" in training data.
Mannered prose gives the model a recognition test: for each sentence, is this phrase displaying the writer or conveying the idea? That is the same design principle behind effective eval rubrics — name the failure mode, give examples, state the fix — which explainx.ai's master prompt-engineering guide treats as the difference between prompts that change behavior and prompts that merely occupy context.
If you are building an agent that writes user-facing copy, consider putting the full paragraph in your system prompt and the short "remove all mannered prose" line in an edit pass after the first draft. Two-pass workflows — generate, then revise against a named style rule — tend to outperform asking for perfect prose in a single shot, especially on long-form content.
Related on explainx.ai
- Master prompt engineering with Claude: complete guide
- Context engineering vs. prompt engineering
- Claude Opus 5's load-bearing "Claudisms" and writing tells
- Claude Fable 5.1 and Mythos 5.1: benchmarks, pricing, and safeguards
- What is CLAUDE.md?
- Top 10 tips for working with Claude efficiently
- Top 10 signs of AI-generated text
- Claude Code token efficiency and prompt cache sessions
- ChatGPT Work writing-style personalization
Sources
- elvis (Elvis Saravia) on X, September 8, 2026
- DAIR.AI Academy — interactive "Write better with Claude Fable 5.1" tutorial
- Anthropic — Prompting Claude Fable 5.1, Writing density section
Prompt text and cross-model results reflect elvis's September 8, 2026 report and Anthropic's official Fable 5.1 prompting documentation. Results from any single prompt vary by task and model version — test on your own writing before adopting it as a standing rule.