Anthropic's tests tie Claude Fable 5.2's code merging to one prompt sentence

single source· 1 articles · confidence: low · first seen 2026-11-10 05:34 UTC

What this means for you

If you want the model to restructure code rather than extend it, say the codebase is its own — Anthropic's suggested wording is one line of prompt. If you want your existing structure kept, say nothing about ownership. These are the vendor's own small-task tests, with no evaluation date and no described harness, so treat the ratio as a prompt hint rather than a measurement.

Anthropic's prompting guide for Claude Fable 5.2 says a sentence about who owns the code changes how much the model restructures it. In Anthropic's testing on small coding tasks, when the prompt said nothing about ownership, or framed the codebase as someone else's, the model tended to add to existing structure even when that repeated code. When the prompt said the codebase was the model's own, it merged repeated code in about half of tries; without the sentence, under a tenth. No evaluation date or harness is given, and the account comes from Simon Willison quoting the documentation.

Models in this story

Key facts

  • ·Anthropic's prompting documentation for Claude Fable 5.2 recommends stating that the codebase belongs to the model when structural improvement is wanted. source
  • ·In Anthropic's testing on small coding tasks, the model merged repeated code in about half of tries with that ownership sentence, and in under a tenth of tries without it. source
  • ·With no ownership statement, or with the code presented as someone else's, the model tended to add to existing structure even when that repeated existing code. source
  • ·The figures are reported without an evaluation date or a description of the test harness. source

What the sources say

  • Simon Willison — Quotes Anthropic's prompting guidance, isolating the ownership sentence and the two duplicate-merging rates.

Sources

The original reporting. Follow these — they did the work.

← the wire