Prompting Claude Fable 5.2

Claude Fable 5.2 requires prompt and effort settings to be retuned rather than inherited from earlier models. The guide recommends starting at high effort, then testing lower settings against task-specific evaluations: medium often suits interactive coding and routine work, while high is useful for difficult or unattended tasks. Moving above high can increase token use substantially for uneven gains. High effort also makes even simple replies slower, and lower effort makes long tasks more prone to stopping early. For agentic work, prompts can explicitly instruct the model to work through ordinary obstacles, continue independent tasks, and stop only for genuine blockers or required safety approvals. Testing found this reduced unfinished sessions, but increased output tokens by about 30 percent.

The guide also explains how instructions affect coding and agent behavior. Saying the model owns and maintains a codebase encourages consolidation of duplicate code - in tests, this rose from under 10 percent of runs to about 75 percent - but can trigger unrequested changes, increase output by 70-80 percent, and alter existing behavior. Approval rules should identify precisely which actions require confirmation, since broad rules can hold up routine work. Long tool-use turns may appear silent because progress updates arrive in thinking blocks that some clients do not display; clients should render them when useful. The guidance further flags setting expectations for subagents, simplifying prompts and harnesses inherited from earlier models, making final reports easier to follow, and handling refusals appropriately.