Anthropic publishes prompt guide for Claude Fable 5.2 developers
The guide targets agent behaviour such as early stopping and sub-agent calls, and reports one instruction that roughly halved premature stops.
Anthropic has published an official prompt guide for its latest model, Claude Fable 5.2, in its developer documentation, according to aitimes.com. The guide addresses problems that show up in real development work, including long tasks that stop early, unnecessary approval requests, the scope of code changes and the use of sub-agents.
One example in the guide concerns a single added instruction stating that the codebase belongs to the model. According to Anthropic, that sentence raised the share of duplicate code found and consolidated from under 10% to about 50%. It also increased output tokens by roughly 1.5 times and led to unrequested test and documentation changes as well as changes to existing behaviour. Anthropic advises against using the instruction for tasks that need only minimal changes.
A second instruction, telling the model not to stop unnecessarily while work remains, roughly halved cases of stopping with completable work left, at a cost of about 30% more output tokens. Specifying the cost and conditions for calling sub-agents cut calls to about one third while keeping task success rates in tests.
Reasoning levels and migration advice
Anthropic said most performance gains appear at the default “high” reasoning setting, and that higher levels may add limited benefit depending on the task. For teams moving to Fable 5.2, the company recommends testing existing prompts on real work and revising instructions one at a time while measuring both performance and token cost.
Prompt guides have increasingly focused on controlling when an AI agent stops, what it handles on its own and how much cost it uses, rather than simply getting a desired answer. OpenAI issued a separate guide when it launched GPT-5 last year, and has recently introduced prompt-writing tips for its image generation model.
Quick answers
What is the Claude Fable 5.2 prompt guide about?
It covers real development problems such as long tasks stopping early, unnecessary approval requests, the scope of code changes and sub-agent use, and gives instructions and measured effects for each.
Which reasoning setting does Anthropic recommend for Fable 5.2?
Anthropic said most performance gains appear at the default “high” setting, and that higher levels may add limited benefit depending on the task.
How should teams move existing prompts to Claude Fable 5.2?
Anthropic recommends testing existing prompts on real work and revising instructions one at a time while measuring performance and token cost.