Anthropic measures what a code-ownership sentence does to Fable 5.2

According to the guide, when a prompt said nothing about ownership or presented the code as someone else's, the model "tended to add to the existing structure, even when that repeated existing code." Anthropic's suggested line is "This codebase is yours: what it's for and how it's built are your call, and you'll keep building on it." On small coding tasks, the model merged repeated code in about half of the tries with that sentence and in under a tenth without it. The sentence has costs. Output tokens rose about 1.5 times, and the model made changes nobody asked for, many of them new tests and documentation. Apart from bug fixes, it changed how an existing feature behaves in one try in six to ten, more often at higher effort. A "No drive-by refactors, renames or cleanups" rule in the same prompt still left merges in about a third of tries.

The guide measures other harness text the same way. An approval rule that listed "sending messages to people outside this conversation" made the model hold back roughly half of its routine requests to colleagues in simulated tasks with nobody there to approve. One stale "never retry" note, among a dozen notes from earlier sessions, cut retries after a busy tool to about four in ten, from nearly all. Text telling the model to finish the whole task cut sessions that ended with doable work left to about half, in replays at medium effort, and raised output tokens by about 30 percent. All of these are Anthropic's figures from its own tests.

Why it matters: The guide treats prompt lines and accumulated agent notes as switches with measured side effects, so a harness carried over from an earlier model needs retesting, not just a new model name. If you add the ownership sentence to get cleanup, review the diff, because Anthropic says the model's closing summary "often left out small ones."