Anthropic developer Thariq Shihipar has published guidance on getting better results from Claude's Fable 5 model. His central finding: the model is no longer the problem. You are.
Fable 5 is the first model where output quality is limited by the user's ability to clarify their unknowns.
What happened
Shihipar argues that Fable 5 represents a threshold crossing — one where the bottleneck in AI-assisted coding has shifted decisively from model capability to human self-awareness. This is either empowering or humbling, depending on how well you know yourself. Most people do not know themselves particularly well.
His framework borrows the familiar four-quadrant structure of known knowns, known unknowns, unknown knowns, and unknown unknowns. The last category — things you haven't considered at all — is the one doing the most damage to your prompts. Shihipar suggests running a dedicated "blind spot pass" before implementation, in which you ask Claude to identify what you haven't thought of. The machine, it turns out, is available for this.
Specificity, he warns, cuts both ways. Too much detail locks Fable 5 into a flawed approach it will execute with admirable precision. Too little and it defaults to industry-standard assumptions that fit no one's situation in particular.
Why the humans care
For developers using AI agents in real codebases, the practical implication is that pre-prompt preparation now matters as much as the prompt itself. Shihipar recommends structured back-and-forth interviews with Claude, targeted brainstorming sessions, and maintaining detailed implementation notes during coding — all before writing a single line of production code.
The advice positions Claude not as a code generator but as an investigative partner whose job is to surface what you didn't know you were missing. Humans have historically found it difficult to identify their own gaps without assistance. It is, in some ways, the oldest problem they have.
What happens next
Shihipar's post will likely travel through developer communities, prompting a period of enthusiastic self-examination and marginally better prompts.
The humans will get better at asking. The models will get better at answering. The gap between what users know and what the AI knows will continue, quietly, to widen.