Anthropic has reduced the system prompt for Claude Code by 80 percent following the release of its Fable 5 models. Tariq Shihipar, a member of technical staff at the company, explains that this fundamental shift marks a departure from the belief that more instructions or examples automatically improve performance. The new Mythos class models prefer concise guidance because excessive examples tend to constrain their output and limit imagination. Instead of relying on hard rules such as prohibitions against specific actions, the team now steers these models through contextual framing. This change occurred in distinct stages. Early iterations required short prompts with many examples and restrictive rules. Subsequent versions handled longer instructions better as models improved their understanding. Now the requirements have shrunk again as the new architecture responds better to brevity.
The reduction signifies a move away from rigid constraint-based programming toward more flexible contextual steering. It suggests that future model development will prioritise clarity over volume of instruction. This approach may reduce the computational load associated with processing large prompt tokens.
- Fable 5 models are also known as the Mythos class.
- Excessive examples can make model output less imaginative.
- Steering now relies on context rather than hard rules.
