Tip: Newer Anthropic Models Do Worse With Custom Edit Tools — Prefer the Native Editor Format

Simon Willison·high signal

Willison documents ("Better Models: Worse Tools", July 4) that Opus 4.8 and Sonnet 5 can perform worse than older versions when driving custom file-edit tools, because they're increasingly trained/optimized for Claude Code's native editor. The takeaway for anyone shipping their own agent: don't invent a bespoke edit-tool schema — mirror the native diff/replace format the model was tuned on, or you'll regress on the newest models.