This pack follows Anthropic's Prompt Guide for Claude Fable 5.2. It contains nine entries: all six complete prompt blocks from the guide, its inline staging-approval example, and two prompts generated based on its recommendations.
The quoted examples retain their original wording, headings, and paragraph breaks. They cover finishing work, code ownership, three ways to keep users informed, and clear final reports. The staging-approval example is a user reply for authorizing a specific action.
The two generated prompts cover delegation and scoped approval rules. Their descriptions and additional information say: “This prompt was generated based on the Prompt Guide for Claude Fable 5.2 by Anthropic.” Each entry links its source section.
Recommendations to try
Start at high; compare medium and xhigh on your evaluations. Audit inherited instructions before adding more. Test long tasks, then address observed problems individually. Measure quality, latency, and token costs together.
Keep approval boundaries specific. Reserve ownership language for projects where broader edits are welcome. For progress visibility, check the client's thinking-block rendering. Keep conversation history append-only; apply system and tool changes in new conversations.
Review the surrounding instructions
Check delegation rules and subagent costs. Trim stale cautions, and state approval exceptions within the rules themselves. Reevaluate inherited token-budget text and rigid style requirements.
For refusals, inspect the API response category. Use supported thinking displays for reasoning summaries.
Evaluate the prompts on your tasks
Review each entry alongside its linked section. Choose a representative task and record a baseline before changing instructions. Keep the task and acceptance criteria consistent across trials. A useful result should tell readers when an instruction helps, what it costs, and when to leave it out.