Anthropic Removes 80% of Claude 5 System Prompts

Anthropic technical staff member Thariq Shihipar announced new context engineering rules for the latest Claude 5 generation models (such as Opus 5 and Fable 5) on July 24, 2026. The core change is the removal of over 80% of the system prompts, with no measurable performance loss in their internal coding evaluations. This decision stems from the assessment that as the model's judgment capabilities improve, previous strong guardrails (rules) actually cause instruction conflicts and waste attention budget.

Paradigm Shift in Context Engineering

In previous models, explicit prohibition rules and examples were essential to prevent worst-case scenarios like file deletion. However, in the latest models, such instructions act as constraints that hinder judgment. Anthropic describes this as 'Unhobbling Claude' and presents the following six inversions of conventional wisdom.

• From Rules to Judgment: Remove strong instructions that are not always true, and guide the model to make its own decisions based on surrounding context.
• From Examples to Interface Design: Instead of tool usage examples, design parameters themselves to be expressive to convey usage.
• From Full Disclosure to Progressive Disclosure: Do not place all information at the beginning; load context via skills or lazy-loading tools when needed.
• From Repetition to Concise Tool Descriptions: Eliminate duplication between system prompts and tool descriptions, leaving guidelines only in the tool descriptions.
• From Manual CLAUDE.md Logging to Automatic Memory: Transition from users manually recording information to Claude automatically learning and saving it across sessions.
• From Simple Specifications to Rich References: Utilize more accurate reference materials such as HTML artifacts, test suites, and rubrics instead of Markdown text.

Managing Attention Budget and Context Rot

LLMs have a finite Attention Budget, and a phenomenon called 'Context Rot' occurs where information recall ability degrades as the number of tokens in the context increases. Therefore, the goal is to find the minimum token set capable of generating the desired results. The drastic reduction of system prompts aims to prevent the model from wasting budget on information included in instructions that it can already replace with judgment.

Practical Application and Tools

Developers can adjust the size of their skills and CLAUDE.md files using the claude doctor command. Additionally, the automatic memory feature is enabled by default and saved in Markdown format at the path ~/.claude/projects/<project>/memory/. It is recommended to keep CLAUDE.md thin, containing only repository purpose and codebase pitfalls, while separating detailed procedures into skills to follow a progressive disclosure structure. However, these rules apply to the latest generation models; excessive rule removal in smaller or previous generation models may lead to failure, so caution is required.