Anthropic: Telling Claude Fable 5.2 the Codebase Is Its Own Raised Code Merging to About Half of Tries

Anthropic's documentation for prompting Claude Fable 5.2 says the model responds closely to what the prompt states about who owns the code it works on. According to Anthropic's testing on small coding tasks, one added sentence changed how often the model merged duplicated code from under a tenth of tries to about half. The passage was highlighted on 10 November 2026 by Simon Willison on his weblog.

Default behavior

Anthropic reports that when a prompt said nothing about ownership, or presented the code base as someone else's, Claude Fable 5.2 tended to add to the existing structure. It did this even when the addition repeated code that already existed.

The recommended sentence

For users who want the model to improve the structure as it works, the documentation advises stating that the code is the model's own. Anthropic's suggested wording:

This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it.

Anthropic's reported results

The figures below come from Anthropic's own testing on small coding tasks, as stated in its documentation:

• With the sentence: the model merged repeated code in about half of the tries.
• Without the sentence: the model merged repeated code in under a tenth of the tries.

The documentation does not give exact counts in the quoted passage, and the figures are Anthropic's own results rather than independent measurements. The testing is described as covering small coding tasks only.

What this means

The guidance shows that, for Claude Fable 5.2, a sentence about ownership is a behavioral lever, not just context. Prompts that say nothing on the subject, or that frame the code as belonging to someone else, correspond to the additive behavior Anthropic describes. Teams that want the model to consolidate duplication as it works have a concrete wording to try.

The reported effect is large in Anthropic's tests, but the scope is narrow: small coding tasks. Anyone applying the sentence to larger repositories should check the results on their own code before relying on the behavior. Because the model also follows the opposite framing, the ownership statement should be chosen deliberately in system prompts and agent instructions, not left to chance.

The passage is a quotation from Anthropic's documentation, and the full guide may contain further recommendations for Claude Fable 5.2. Pricing, context window and benchmark figures are not part of this passage.