The updated prompting guide for Claude Fable 5.2 emphasizes that this model follows instructions very closely, taking users at their word. Unlike older models, pushing it harder is less effective; instead, clarity and precision in prompts are key. The guide, based largely on the company’s internal testing, suggests starting with a high effort setting on the API for most tasks, as this balances performance and token usage. Lower effort settings are better for quick back-and-forth interactions, while going above high yields uneven gains and higher token consumption. For straightforward questions, answering directly without over-deliberation is recommended, and if responses feel slow, lowering the effort dial can help.
One common issue with long tasks is that Claude sometimes stops prematurely, especially at low and medium effort levels. The guide advises diagnosing ordinary obstacles like errors or timeouts and working through them, while deliberate blockers should be left alone and reported. Using medium effort reduces the frequency of unfinished work but increases token usage by about 30%. Because the model asks for confirmation less often at higher effort, users should maintain their own checks for risky operations. The guide also stresses the importance of clear communication, such as explicitly stating approvals rather than using vague confirmations like thumbs up or “continue.”
Ownership and clarity in prompts are crucial for effective code generation. When ownership of the code is unspecified, Claude tends to add redundant code and report issues without fixing them. Simply stating that the codebase belongs to the user encourages the model to merge repeated code and produce cleaner outputs, though this can increase token usage and sometimes introduce unwanted changes. Users should carefully review code changes rather than relying solely on summaries. Additionally, outdated cautionary notes or instructions can cause the model to behave overly cautiously, holding back routine work unnecessarily. Updating or removing such legacy instructions and being specific about exceptions improves performance.
The guide recommends asking for periodic updates during long sequences of tool calls to prevent the model from going silent. Providing one-line progress updates every few calls keeps users informed, and final reports should be tailored for readers unfamiliar with the work, using clear language and avoiding jargon or invented abbreviations. Delegation of tasks to helper models should be managed carefully; removing unnecessary delegation instructions and clarifying the cost and value of helpers can reduce excessive delegation without harming task success. Builders are advised not to request the model to write out its reasoning in replies, as safety classifiers may block such requests—summarized thinking is preferred.
To implement these changes, the guide suggests starting with the high effort setting, rewriting all carried-over prompts to reflect current needs, and running the longest tasks to identify where the model stops or goes quiet. Fixing issues one at a time and measuring improvements is recommended. The overarching message is that Claude Fable 5.2 is a powerful tool when given clear, current instructions—users should say exactly what they mean and remove outdated or irrelevant directions. For those seeking to improve their AI workflows, the presenter offers a course called Get Amplified, live weekly calls, and consulting services to help users adapt and thrive with evolving AI capabilities.
Useful Links
