In the rapidly evolving world of artificial intelligence, a quiet but significant revolution is taking place. The industry leader, Anthropic, has made a seemingly paradoxical move: it removed more than 80% of the system prompt for its flagship tool, Claude Code. In an era when everyone is chasing increased complexity and data volume, this move demonstrates technological maturity and a paradigm shift in the approach to interacting with AI.
Why less became better
My analysis of the internal changes shows that the cause was a fundamental contradiction in the old instructions. Engineers discovered that the numerous rules, skills, and requirements in the prompt conflicted with one another, creating "cognitive dissonance" for the model. One part of the context demanded one thing, another demanded the exact opposite. This led to instability and a decline in response quality.
The key takeaway I draw from Anthropic's new strategy is that next-generation models no longer need rigid oversight. They are capable of independently analyzing the surrounding context and making decisions. Instead of trying to anticipate every step, developers now trust the model. And, as internal tests show, this decision not only did not worsen but actually improved performance in programming tasks.
New rules of context engineering
Instead of bulky instructions, Anthropic proposes a new, minimalist approach. The main principles I highlight from the new paradigm are:
- Trust in the model's judgment. Instead of prohibitions, there is one simple recommendation: write code in the style of the surrounding project.
- Avoiding excessive examples. Three demonstrated ways to solve a task will be perceived by the model as the only possible ones, limiting its creativity.
- Gradual information disclosure. There is no need to dump all instructions into one file. The model will pull in the necessary data at the right moment on its own.
- Eliminating duplication. An instruction should be stated once, not repeated in different places, creating confusion.
I also note the shift in emphasis from simple markdown files to more structured formats, such as HTML and artifacts, as well as the use of automatic memory. This allows the model to manage information more efficiently and not waste tokens on processing unnecessary content.
My verdict as an analyst: this move is not just optimization but a signal that AI is transitioning to a new level of autonomy. For developers, this means a paradigm shift: from "prompt engineering" (how to get the model to do something) to "context engineering" (how to give the model the freedom to do it right). Tools that fail to adapt to this philosophy risk being left behind.