Anthropic has radically rethought its approach to context preparation for the new generation of Claude models. During an internal audit of Claude Code logs, it turned out that more than 80% of the system prompt was dead weight: instructions contradicted each other, duplicated, and only confused the model. After the cleanup, the quality of work did not drop—on the contrary, it improved.
This is about so-called context engineering—everything the model reads before the user's request: the system prompt, skills, project files, and memory. It is here that Anthropic discovered a systemic problem: the system prompt, skills, and user request conflicted with each other. One part of the context demanded one thing, another—the exact opposite.
Why the old rules are outdated
The new generation of Claude models has learned to independently read the surrounding context and make decisions without needing rigid prescriptions. Therefore, the company abandoned most of the old rules. At the same time, in internal programming tests, quality remained at the same level. The essence of the changes is simple: Claude works better for the first time with fewer instructions, and the user should not overload the model but give it more freedom.
The new rules look like this. First, trust the model's judgment: instead of prohibiting, for example, writing comments in code, one line is enough—"write code in the style of the surrounding code." Second, examples are no longer the main tool: three demonstrated methods will be perceived by the model as the only possible ones. Third, gradual disclosure: there is no need to cram all instructions into one huge file; Claude will pull up the necessary information at the right moment itself. Fourth, say it once: duplication of prescriptions is now prohibited—the same rule should not appear in both the prompt and the tool description.
How to bring order to settings
Anthropic's main advice is reduction. Everything the model can understand on its own should be removed from the context. Check each instruction against three questions: can Claude handle it without this, is it already present elsewhere, and is it only needed occasionally—if so, it is better to move it into a separate skill that will be loaded only when necessary.
The system prompt, however, is not important to everyone. An ordinary Claude Code user does not need it, but when creating your own agent, it is precisely where you should focus. The conclusion is simple: context needs to be compressed, leaving only critical instructions, and otherwise—give the model more freedom.
My view: this step is a marker of LLM maturity. We are moving from the era of "prompt engineering" to "context engineering," where quality is determined not by the volume of prescriptions but by the model's ability for autonomous judgment. For developers, this means a paradigm shift: not "make the model do it," but "create the conditions so it does it correctly on its own."