Anthropic has radically rethought its approach to context engineering for the new generation of Claude models. During an internal audit, engineers discovered that the Claude Code system prompt had ballooned to critical proportions, and its instructions contradicted each other. The solution turned out to be unexpected: removing more than 80% of the prompt led to an increase in quality, not a decline.
This is about a fundamental shift in how the model perceives its environment before a request. Previously, the context included the system prompt, skills, project files, and memory—a unified but cumbersome structure. An analysis of Claude Code logs showed that parts of this system conflicted: one part demanded one thing, another demanded the exact opposite.
Why less turned out to be better
The new generation of Claude models, as it turned out, is capable of independently interpreting context and making decisions without rigid prescriptions. Therefore, Anthropic took a bold step—reducing the system prompt by more than 80%. Internal programming tests confirmed: quality not only did not drop, but in a number of scenarios it even improved.
The key takeaway from the company's new guide: the model performs better with a smaller volume of instructions. The user should not overload the system but give it more freedom. Instead of multi-page rules—one line: "write code in the style of the surrounding project." Instead of dozens of examples—one, so as not to limit variability. Instructions that are needed only occasionally are now moved into separate skills and loaded only as necessary.
Anthropic pays special attention to automatic memory and advanced links. Claude saves important information on its own, and instead of simple markdown files, it recommends using HTML and artifacts. The system prompt is not important for everyone: an average Claude Code user does not need it, but when creating your own agent, it is precisely what you should focus on.
My view: This step is a marker of the maturity of the AI industry. We are moving from the era of "prompt engineering" to the era of "context engineering," where the model takes on interpretation, and the human only sets the direction. For developers, this is a signal: the simpler and cleaner the context, the higher the return from the model. Investors and analysts should keep an eye on this trend—it directly affects the performance of AI agents and, consequently, their commercial value.