// AI NEWS · PROMPTING

Anthropic Publishes "Power Prompts" for Claude Fable 5.1 — Copy-Paste Lines That Fix AI-Sounding Writing and Half-Finished Work

Anthropic's official Fable 5.1 prompting guide documents a set of drop-in instructions that fix the most common output complaints — flowery "mannered prose," half-done tasks, and answering from stale memory — most of them short enough to paste into a single message.

2026-09-15 · by Moe Ameen

What happened

Anthropic maintains an official prompting guide for its Claude Fable 5.1 and Mythos 5.1 models that documents the model's behavioral differences from Fable 5 and gives drop-in prompt patterns to correct them. In coverage published on September 15, 2026, Search Engine Journal characterized the highlights as "power prompts" — a phrase Anthropic itself doesn't use, but a fair description of a set of copy-paste instructions that fix the output problems people complain about most. Fable 5.1 shipped on September 1, 2026, and the guidance accompanies it.

The one most relevant to anyone writing content is aimed at flowery, self-regarding writing, which Anthropic calls "mannered prose." Its definition, meant to be pasted into a prompt, is precise: mannered prose "substitutes metaphor and flourish for direct statement. Instead of 'a parameter worth varying,' the mannered writer produces 'a dial worth turning.'" The fix Anthropic offers is to define the anti-pattern and then say what you want instead — and it notes that the short version, "Please remove all mannered prose," also tends to work. A companion pattern addresses formatting: because Fable 5.1 reaches for bold, headers, and lists less than earlier models, Anthropic suggests removing old anti-formatting rules and instead telling the model when specific formatting is appropriate.

Several other patterns target agentic and research behavior rather than prose. "Finish the whole task" is a system-prompt block that tells the model it is operating autonomously and should carry out reversible next steps instead of stopping to ask "Shall I…?" or describing what it would do next. "Search triggering at low effort" nudges the model to verify unfamiliar names — especially in fast-moving areas like AI tools — by searching the name as the user wrote it rather than answering from memory. Others cover preferring targeted edits over full-file rewrites, keeping changes and tests scoped to what was asked, and reducing safety-classifier false positives on benign code.

Two caveats worth keeping in view: the guidance is tuned to Fable 5.1's specific tendencies, so a pattern written for it isn't automatically the right instruction for a different model, and the "power prompts" label is a secondary framing, not an Anthropic product. The prompts themselves, and the mannered-prose definition, are quoted from Anthropic's own documentation.

Why it matters for creators

  • The mannered-prose prompt is the most direct fix yet for the "AI tell" that gets flagged writing demoted and distrusted. Naming the anti-pattern and asking for plain statement instead is a repeatable lever, not a vibe.
  • These are instructions, not a feature — they help one chat reply at a time. The gap between "a cleaner draft" and "a week of on-brand posts published everywhere" is still yours to close.
  • The "finish the whole task" and "search the name as written" patterns matter for creators who run AI as an assistant: they cut the back-and-forth and reduce confidently wrong claims about new tools — exactly the errors that sink a first-mover post.
  • Prompt patterns are model-specific. A line tuned to Fable 5.1's density won't behave the same on an older Claude or another vendor's model, so treat any pasted prompt as something to test, not a universal setting.
  • The signal under all of it: getting good output is increasingly about the standing instructions you attach, not the single clever request — which is an argument for encoding your rules once rather than retyping them.

How to act on this with Kompozy

The useful way to read this isn't "here are prompts to paste" — it's that Anthropic just published, in its own words, the exact writing discipline that separates content from AI slop: kill the mannered prose, say what you mean, hold the format to what the piece needs. The catch is that a pasted prompt applies that discipline to one reply, in one chat window, that you then still have to shape and post yourself. [Kompozy](/) exists to make that discipline a standing property of everything you ship rather than a line you remember to include. The angle here is governance, not generation: the same rules Anthropic hands you as a prompt, Kompozy enforces automatically on every output.

Concretely, the [Persona Brief](/glossary/persona-brief) is where you set the "power prompt" once — your voice, your point of view, and a banned-word list that includes the mannered-prose flourishes and stock AI phrasing you never want to see — and it then governs every generation, not just the next one. From a single source, Kompozy produces the whole set under that brief: platform-fit Text Posts, brand-exact Carousels and Quote Graphics via [HyperFrames](/glossary/hyperframes), avatar-narrated [Persona Shorts](/glossary/persona-shorts), [Clipped Shorts](/glossary/clipped-short), a blog article, and an email newsletter — each already written to the standard, then scheduled and fanned across the eight social platforms plus blog and email by [Autopilot](/glossary/autopilot) behind a per-post review gate. Anthropic gave you the instruction; Kompozy is where you stop retyping it and let it run at the scale of a real publishing schedule. To build the governing instruction itself, start with [how to write a system prompt](/how-to/write-a-system-prompt).

Quick takeaways

  • Anthropic's official Claude Fable 5.1 prompting guide documents drop-in prompt patterns; coverage on September 15, 2026 dubbed the highlights "power prompts." Fable 5.1 launched September 1, 2026.
  • The headline pattern targets "mannered prose" — flowery writing that substitutes metaphor for plain statement (its example: "a dial worth turning" vs. "a parameter worth varying"). The short fix, "Please remove all mannered prose," also works.
  • Other patterns cover formatting-on-demand, finishing a whole task without asking permission mid-way, verifying unfamiliar names by searching, targeted edits over rewrites, and scoping changes to what was asked.
  • The patterns are tuned to Fable 5.1 specifically — test before assuming they transfer to another model — and "power prompts" is a secondary label, not an Anthropic product name.
  • The deeper point for creators: quality follows the standing instruction, not the one-off request — which is what an engine like Kompozy encodes once via a Persona Brief and applies to every published piece.

Frequently asked questions

What are Anthropic's "power prompts"?

"Power prompts" is a secondary label for a set of copy-paste prompt patterns in Anthropic's official Claude Fable 5.1 prompting guide. They fix common output problems: flowery "mannered prose," half-finished agentic tasks, answering from stale memory, and over-formatting or under-formatting. Anthropic doesn't use the phrase itself; a September 15, 2026 recap coined it.

What is the "mannered prose" prompt and does it fix AI-sounding writing?

It defines the anti-pattern for the model: mannered prose substitutes metaphor and flourish for direct statement — writing "a dial worth turning" instead of "a parameter worth varying." Pasting that definition (or the short "Please remove all mannered prose") pushes the model toward plainer, more direct writing, which is the most common single fix for the AI-tell that gets content flagged.

Do these prompts work on any AI model?

No — they're tuned to Claude Fable 5.1's specific tendencies versus Fable 5. A pattern that corrects Fable 5.1's density or formatting may do nothing, or the wrong thing, on an older Claude model or a different vendor's model. Treat any pasted prompt as something to test on your own task rather than a universal setting.

How do I apply this discipline across all my content, not just one reply?

A pasted prompt governs a single response. To make the same rules — plain voice, banned AI-tell phrases, format discipline — apply to every post, you encode them once as a standing instruction. Kompozy does this through a Persona Brief that governs every generated Text Post, carousel, short, blog, and newsletter, then publishes the set across the eight social platforms plus blog and email.

Related news

← All AI news · Get started →