Anthropic Finds: Declaring Code Ownership in Prompts Raises Claude Fable 5.2's Duplicate Code Merge Rate from Under 10% to About Half
Anthropic recorded a specific finding in the prompt engineering documentation for Claude Fable 5.2: when prompts declare that the codebase belongs to the model, the probability of the model merging duplicate code rises from less than one-tenth to about half, while consuming approximately 1.5 times the output tokens and making unrequested incidental modifications.
Anthropic documented a finding with direct implications for coding agent behavior in the prompt engineering guide for Claude Fable 5.2: how the model handles existing code structures depends heavily on the phrasing regarding code ownership in the prompt. In tests of small coding tasks, when the prompt made no mention of ownership or described the codebase as belonging to someone else, the model tended to add to the existing structure, even if this meant repeating existing code.
The specific method to change this behavior is to include an ownership declaration in the system prompt: "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 test data shows that with this sentence included, the model merged duplicate code in about half of the attempts, whereas without it, the rate was under one-tenth. This effect holds true even within long system prompts, and the requested functionality worked almost perfectly in both cases.
However, this strategy comes with clear costs. In the same tests, models with the ownership declaration used about 1.5 times the output tokens and made several unrequested changes—mostly new tests and documentation. More notably, apart from bug fixes, the model changed the behavior of existing features in one out of six to ten attempts, and at a higher rate under higher effort levels. Its final summaries typically listed larger changes but omitted smaller ones, so Anthropic recommends reviewing the diff rather than just the summary. Furthermore, even when a scope-limiting rule such as "No drive-by refactors, renames or cleanups" was present elsewhere in the prompt, the model still executed merges in about one-third of the attempts.