Fable 5 and Fable 5.1 Prompting Guide

Build prompts that give Claude Fable a clear reason, outcome, boundary, and evidence bar while leaving room for strong long-horizon judgment. Remove legacy micromanagement before adding model-specific scaffolding. Fable 5 prompts carry over to Fable 5.1 unchanged; the shifts section below covers the behaviors that moved.

Quick Start

1. Identify the target artifact and workload: new prompt, review, migration, tool contract, long-running agent, or failure diagnosis. Name the exact model, Fable 5 or Fable 5.1, and the runtime it runs in.
2. Define the observable outcome, why it matters, relevant context, authority boundary, evidence, output, and completion criteria.
3. For an existing prompt, inspect available traces or checks. Separate an observed failure from a suspected prompt cause; runtime availability alone does not authorize comparison runs.
4. Remove repeated rules, exhaustive behavior lists, stale thinking-budget instructions, aggressive tool triggers, anti-formatting and hold-your-findings rules written for chattier models, and scaffolding for behavior Fable already performs reliably.
5. Add brief, direct clauses only for real gaps. Use references/prompt-patterns.md for long-run, tool, memory, and communication patterns. Use assets/prompt.template.md for a new complex prompt.
6. Put effort, thinking display, timeout, fallback, compaction, history handling, and tool setup in the API or harness rather than pretending they are prompt instructions.
7. Validate the requested artifact. When model evaluation is requested, test the same cases at the intended effort and compare outcome correctness, evidence, scope control, latency, tokens, and cost.
8. Deliver the prompt, required runtime settings, and concise rationale. Request evidence and decisions, never a transcript of private reasoning.

Read references/source-notes.md when maintaining model-specific claims or API behavior.

Core Prompt Shape