Anthropic removes 80%+ of Claude Code's system prompt for Claude 5
Original: The new rules of context engineering for Claude 5 generation models
Why This Matters
Signals a shift in LLM deployment: more capable models need less hand-holding, reshaping agent architecture norms.
Anthropic announced it removed over 80% of Claude Code's system prompt for Claude 5-generation models like Claude Opus 5 and Claude Fable 5, with no measurable loss on coding evaluations. The company shared updated context engineering best practices for developers building agents.
Anthropic published new guidelines for context engineering targeting Claude 5-generation models, revealing that the company removed over 80% of Claude Code's system prompt when deploying models such as Claude Opus 5 and Claude Fable 5 — without any measurable degradation on coding benchmarks.
According to the post, context engineering differs from traditional prompting because context is applied broadly across many requests rather than tailored to individual queries. Sources include system prompts, Skills, CLAUDE.md files, and memory. Anthropic noted that the previous system prompt was 'overconstraining' Claude, limiting its improved native capabilities.
The key finding is that newer Claude 5 models require significantly less explicit instruction to perform well. Anthropic describes the shift as 'unhobbling Claude' — reducing prescriptive guidance and allowing the model's built-in reasoning to operate more freely.
For developers, Anthropic introduced a `/doctor` command in Claude Code (invoked via `claude doctor`) that automatically rightsizes Skills and CLAUDE.md files to align with these updated best practices. The guidance is intended to help both Claude Code users and developers building their own agent systems to avoid over-engineering context for the new model generation.