XNews

Anthropic Slashes System Prompts 80%: From Rules to Judgment, Skills, Auto-Memory

Anthropic Slashes System Prompts 80%: From Rules to Judgment, Skills, Auto-Memory

With Fable 5 and Opus 5, Anthropic cut system prompt volume by 80% without losing coding benchmark performance by trusting the model's judgment, shifting to interface-based guidance, progressive Skill disclosure, and automatic memory. Key shifts: allow models to infer style from surrounding code rather than explicit rules; use parameterized interfaces instead of fixed examples; split independent information into on-demand Skills; remove repetitive warnings between system prompts and tool descriptions; replace manual claude.md entries with auto-memory; and use rich references like images, HTML, test code, and scoring rubrics as context. Practical recommendations include keeping claude.md minimal, adopting modular Skills, and using the `claude doctor` command to automatically trim instructions.

Source posts