What's this about?
Simon Willison published an annotated interview with Anthropic's Cat Wu and Thariq revealing that Claude Code's system prompt shrank ~80%: with capable new models like Fable, you should drop the piles of examples and "don't do X" lists that older models needed. Mollick adds that Opus-era "skills" can actively *hurt* Fable's outputs.
What exactly is a "system prompt"?
The hidden standing instructions a product gives the model on every request — its rulebook and persona. Teams used to pack these with defensive examples and prohibitions to steer weaker models. The news: smarter models follow intent better, so heavy scaffolding now adds noise and can degrade results.
Why does it matter?
Practical guidance for anyone building on Fable: prompts and skills tuned for older Claude models may need to be *cut down*, not carried forward — less is more with the new generation.