Anthropic cuts Claude Code system prompt 80% as newer models need less instruction

Key takeaway

• Anthropic has cut the system prompt for Claude Code by 80 percent, reflecting a fundamental shift in how newer Fable 5 models respond to instructions.

• Rather than requiring lengthy, detailed rules and examples, these newer models perform better with shorter prompts and work more creatively within broader context-based guidance.

• This suggests that as AI models become more capable, the way companies steer and control them may need to change dramatically.

3 Key Points

FAQ

How has the approach to instructing AI models changed at Anthropic?

Early models needed short prompts with lots of examples and restrictive instructions. Prompts then grew longer as models improved at understanding them. With the newer Fable 5 models, prompts are now getting shorter again, and Anthropic is steering them through context rather than hard rules like 'do not do this.'

Why do the newer Fable models perform better with smaller prompts?

According to Tariq Shihipar, examples 'tend to constrain it because it's actually more imaginative than the examples we give it.' The models are capable enough that extensive examples and detailed rules can actually limit their performance rather than improve it.