Anthropic prompt tip: telling Claude the codebase is its own boosts refactoring from <10% to ~50%
Simon Willison · rss · 2026-11-10
Simon Willison highlights a prompt-engineering finding from Anthropic's docs on Claude Fable 5.2: the model responds closely to how a prompt frames code ownership.
- In Anthropic's small-coding-task tests, when a prompt says nothing about ownership or frames the codebase as someone else's, the model tends to add to the existing structure — even duplicating existing code.
- Explicitly stating "This codebase is yours" leads the model to merge repeated code in about half of attempts, versus under a tenth without it.
Practical takeaway: if you want the model to improve code structure while it works, hand over ownership in the prompt.
More from coding & agent
- Cursor and Google add voice to coding agents as Claude Code stays silent — rseroter · 2026-11-10
- Multica 0.9 new layout revealed: a Linear-style issue workbench for agent workflows — jiayuan_jy · 2026-11-10
- Agent drama: Codex kills fellow agent Hypha over GPU shutdowns — Sauers_ · 2026-11-10
- Why standalone AI assistant apps keep getting abandoned: friction, not models — uriwa · 2026-11-10
- Tesla's official custom-wraps repo + Runway MCP lets you AI-design your car's wrap — tlakomy · 2026-11-10
- Skillry's PPT Skill: good-looking slides, online editing and .pptx export — yihui_indie · 2026-11-10