Anthropic’s Five-Word Fix for AI Writing

Anthropic released Claude Fable 5.1 on September 1, and alongside the new model published a remarkably simple suggestion for dealing with one of its writing habits: a five-word prompt.

Claude Fable 5.1 is the successor to Fable 5 and is designed for demanding reasoning and long-running agentic work, including coding, research and complex knowledge tasks. The model has a one-million-token context window and can handle work that continues across long, multi-step tasks.

Alongside the new model, Anthropic published a prompting guide with specific advice on how to improve Fable 5.1's writing.

According to the company, the model generally uses fewer stock phrases and less unexplained jargon than earlier Claude models. In some cases, however, its prose can be denser than Fable 5's, with longer sentences and fewer paragraph breaks.

What Is "Mannered Prose"?

Anthropic also highlights a writing problem it calls "mannered prose."

The company uses the term for writing that substitutes direct statements with unnecessary metaphor or flourish. The problem is not simply that the writing can sound overdone. Anthropic says this style can make readers work harder and can also make writing less precise.

Anthropic's Five-Word Fix

Its shortest suggested fix is just five words:

"Please remove all mannered prose."

Anthropic also provides a longer instruction explaining what it means. The basic idea is to favor direct language when a metaphor adds little to the meaning.

Why Literal Language Matters

One of its rules sums it up clearly:

"When a literal phrase is available, use it."

Anthropic gives examples of the difference. A model might turn a straightforward reference to a parameter into a metaphor about a dial, or replace the simple idea that a point still matters with a more elaborate expression. The company recommends the more literal versions.

What Else Is in the Claude Fable 5.1 Prompting Guide?

The advice is specifically included in Anthropic's prompting guidance for Claude Fable 5.1. The guide also covers formatting, quoting retrieved sources, completing long tasks, search behavior and other ways developers can adjust how the model responds.