Anthropic Publishes an Official Prompting Guide for Claude Fable 5.2: Saying "This Codebase Is Yours" Lifts Duplicate-Code Merging From Under a Tenth of Tries to About Half, and Old Prompts Need a Cleanup

Core Background and Industry Pain Points

After Claude Fable 5.2 launched on November 9, many teams carried over system prompts, skills, and harnesses tuned for Fable 5.1 or Opus 5.5, and hit problems: turns that end before the work is done, long silent agentic runs, too many subagents, and needless approval requests. Anthropic responded with a dedicated Fable 5.2 prompting guide in its developer docs; Simon Willison's post recommending it drew more than 130 bookmarks.

Architecture Highlights and Underlying Mechanisms

The guide is organized by symptom. Fable 5.2 reacts strongly to what the prompt says about who owns the code: with no statement, or when the code is someone else's, it tends to add to the existing structure and report nearby problems without fixing them; told the codebase is its own, it restructures. It also follows cautious rules that pile up in memory and notes, so Anthropic advises deleting stale notes and naming steps that need no approval. Fable 5.2 uses preserved thinking, so editing history invalidates thinking blocks.

Authoritative Benchmarks and Measured Results

All figures are Anthropic's own testing (vendor figures) and have not been reproduced by third parties: the ownership sentence raised duplicate-code merging to about half of tries from under a tenth, at about 1.5x output tokens; the "Finishing work" text cut sessions ending with doable work left by about half, at about 30% more output tokens; opening and recap text made the model speak before its first tool call every time, at about 20% more tokens; and a new subagent tool description cut delegation by about a third with the same task success.

Developer Playbook and Getting Started

Anthropic's suggested migration order: set effort to high explicitly, then also run evals at medium and xhigh; read the system prompts, skills, memory, and notes you carry over and remove "think less" and standing delegation lines; run your longest agentic tasks and note early stops, silence, and over-delegation; then apply the matching guidance one change at a time and measure its token cost. Leave out the ownership sentence when users want minimal changes, and review the diff rather than only the summary.