Anthropic has quietly revolutionized how we interact with its latest flagship models, revealing that it removed over 80% of the system prompt for Claude Code when switching to Claude Opus 5 and Claude Fable 5. This massive reduction in explicit instruction resulted in no measurable loss on coding evaluations, signaling a fundamental shift in how large language models consume context. The announcement, detailed in a blog post by Anthropic technical staff member Thariq Shihipar, challenges the prevailing wisdom of 'context engineering' which has long prioritized verbose, rule-heavy system prompts to prevent model hallucinations and errors.

The Myth of Over-Constraining

For years, developers operated under the assumption that LLMs needed strict guardrails to avoid worst-case scenarios, such as deleting files or writing excessive documentation. Anthropic’s internal data suggests this approach often backfires with newer, more capable models. When the system prompt, skills, and user requests clash—for example, when a system prompt says 'DO NOT add comments' while a user request implies the need for documentation—the model must expend cognitive resources resolving the conflict. Shihipar notes that while older models required these hard constraints to maintain quality, Claude 5 generation models possess enough judgment to interpret user intent from surrounding context, making rigid rules obsolete.

From Rules to Judgement

The core of Anthropic’s new philosophy is moving from 'Give Claude rules' to 'Let Claude use judgement.' A prime example is the shift in code comment guidance. Previous iterations of Claude Code enforced a strict 'no comments' or 'one short line max' rule to prevent the model from generating incorrect or verbose documentation. The new approach replaces this with a contextual directive: 'Write code that reads like the surrounding code: match its comment density, naming, and idiom.' This allows the model to adapt to the specific style of the repository rather than forcing a universal standard that might be inappropriate for complex, multi-line logic blocks.

Progressive Disclosure Over Upfront Loading

Another significant change involves how context is loaded. The era of stuffing every possible instruction into the upfront system prompt is ending. Anthropic is now championing 'progressive disclosure,' where tools and instructions are loaded only when needed. This is achieved through new features like 'deferred loading,' where agents must search for tool definitions using ToolSearch before execution, and by moving specific workflows like code review and verification into separate 'skills' that Claude Code calls selectively. This reduces token bloat and allows for a more dynamic, efficient context window.

Key Takeaways

  • Delete the Manual: Anthropic removed 80% of Claude Code’s system prompt for Opus 5 and Fable 5 without performance degradation.
  • Trust the Model: Newer models interpret intent better; rigid rules like 'no comments' can actually hinder performance in nuanced scenarios.
  • Use 'claude doctor': A new command in Claude Code helps developers automatically rightsize their skills and CLAUDE.md files to match these new best practices.
  • Reference Code, Not Just Text: Specs and references should ideally be in code (e.g., test suites, HTML artifacts) rather than simple markdown descriptions, as the model understands code fidelity better.

The Bottom Line

If you are still writing 2,000-word system prompts to micromanage Claude, you are likely fighting the model rather than working with it. Anthropic’s move proves that for Opus 5 and Fable 5, less is not just more—it is necessary to unlock the model’s true reasoning capabilities.