Anthropic has radically rethought its approach to context preparation for the new generation of its Claude models. During an analysis of internal Claude Code logs, engineers discovered that the system prompt, skills, and user requests were in direct conflict with each other. This led to a decline in response quality and unjustified token consumption. The solution turned out to be paradoxical: the company removed more than 80% of the system prompt, and the quality of model performance on internal programming tests not only did not drop, but actually improved.

What is the essence of the changes

The new Claude models have learned to independently read the surrounding context and make decisions, so the need for rigid, voluminous instructions has disappeared. Previously, developers spent enormous effort creating detailed prompts, trying to anticipate every detail. Now, as my analysis shows, this is not just unnecessary, but also harmful — excessive rules limit the model's flexibility and lead to conflicts.

The key changes in context engineering rules are as follows:

  • Trust the model's judgment. Instead of prohibitions, for example, on writing comments in code, only one line is now recommended: "write code in the style of the surrounding code."
  • Minimum examples. Previously, examples were the main rule, but now they, on the contrary, limit the model, forcing it to perceive the shown methods as the only possible ones.
  • Gradual disclosure. There is no need to cram all instructions into one huge file. Claude will itself pull up the necessary information at the right moment, and extra rules only waste tokens.
  • Say it only once. Duplicating instructions in the prompt and tool descriptions is now prohibited — it only creates confusion.

What this means for users

The main advice from Anthropic is to cut down. Remove everything the model can understand on its own. Instructions needed only occasionally are better placed in separate skills so they are loaded only when necessary. The system prompt is not important for everyone: an ordinary Claude Code user may not need it at all, but when creating your own agent, it is precisely where you should focus.

This step marks a paradigm shift: we are moving from the era of "prompt engineering" to the era of "context engineering," where quality is determined not by the volume of instructions, but by the ability to give the model freedom and the right environment. In my view, this is a logical stage of evolution, and other market players should take a close look at this experience — the future lies in minimalism in settings.