Anthropic developer Thariq Shihipar emphasizes that the effectiveness of outputs from Fable 5 hinges on users' ability to recognize their own knowledge gaps and blind spots before writing prompts. Shihipar outlines several techniques to uncover these 'unknowns,' including targeted brainstorming, structured interviews with the AI, and maintaining detailed notes during coding sessions. He notes that the central challenge in prompting is finding the right level of specificity, as overly detailed instructions can lead to flawed approaches, while overly vague prompts often result in generic responses. Source: thedecoder
Shihipar explains that Fable 5 is the first model where output quality is limited by the user's ability to clarify their 'unknowns.' He categorizes these into 'Known Knowns,' 'Known Unknowns,' 'Unknown Knowns,' and 'Unknown Unknowns,' with the latter being the most critical. According to Shihipar, being too specific or too vague can both hinder performance, as Fable 5 may rigidly follow instructions or rely on industry defaults that don't fit the specific task. He advises users to provide context about their starting point, including their level of experience with the problem, to help the model better understand their needs. Source: thedecoder
Before implementation, Shihipar recommends a 'blindspot pass' where Claude identifies unknown unknowns, especially when working in unfamiliar code areas. He suggests prompts like asking for help with unknowns related to a new auth provider. For areas with many 'unknown knowns,' such as visual design, he recommends brainstorming and prototyping with Claude to generate different design directions. Shihipar also highlights the importance of references, such as source code, and having Claude create an implementation plan that focuses on parts likely to change. Source: thedecoder
