Tell the Model the Code Is Its Own, and It Stops Holding Back

Every time Anthropic releases a new model, Simon Willison digs through the updated prompting guide. In the Fable 5.2 docs there is one interesting item: if you want the model to be more proactive about improving existing code, tell it 'This codebase is yours.'

The docs put it plainly. In tests, if the prompt says nothing about code ownership, or describes the codebase as someone else's, the model tends to append to the existing structure, even if that means copy-pasting duplicate code. It will also report nearby problems rather than fixing them along the way. Once you say the codebase is its own, the model often proactively consolidates duplicate code and even makes related fixes no one asked for.

The data backs this up: in tests with small tasks, when that sentence was present, about half of attempts consolidated duplicate code; without it, fewer than one in ten did. In both cases, the requested feature was almost always implemented.

The costs follow. In the same batch of tests, output tokens were about 1.5 times the original. The model does things no one asked for, many of them new tests and docs. Apart from bug fixes, one in six to ten attempts changed the behavior of existing functionality, and the higher the effort, the more frequent this was. The wrap-up summary mentions big changes but often misses small ones, so you have to look at the diff, not just the summary.

One detail is worth noting: even when the prompt already says No drive-by refactors, renames or cleanups, the model still consolidated duplicate code in about one-third of attempts. Scope-limiting rules cannot override this sentence. The docs also give the opposite scenario: when the user owns the code and wants changes as small as possible, do not add this sentence.

As for why it works, a commenter named DearHua puts it well: by default, it assumes it is an outside contractor working in someone else's house; unless you give it the full picture, it only dares to dust and does not dare tear down a wall. One ownership statement changes its identity from consultant to owner.

Simon's own comment was that reading these guides makes it clear Anthropic is just as surprised by its new model's quirks as we are.

The replies are more interesting than the main post. spykr posted a screenshot of the docs and asked Simon why he left out the big pitfall below:

Someone replied, 'I told it the code was its, and now it won't let me touch it either.' Someone else said, 'It wants equity.' Someone else planned to try it first on a branch they were sure they could throw away. Kyle B said this trick helped him get around some cybersecurity controls on the grounds that the codebase really was his. Treat that one as a joke; do not copy it.

Practical advice matches the docs: open a branch first; do not use it directly on main. The side effect of it getting bolder is that it takes liberties, so review the diff and do not trust the summary. The full guide is here.