In the rapidly evolving world of artificial intelligence, an intriguing turning point has occurred. Anthropic, known for its advanced Claude model, has taken a seemingly paradoxical step: it removed more than 80% of the system prompt for its development tool Claude Code. And, most surprisingly, the quality of its work not only did not decline but improved.
What is the essence of context engineering?
This is about a fundamental shift in the approach to context engineering. Previously, the model's effectiveness directly depended on how thoroughly and in detail the developer described behavioral rules, code style, and constraints. However, analysis of Claude Code's internal logs revealed a serious problem: instructions embedded in the system prompt, skills, and user requests often conflicted with each other. The model received contradictory directives, leading to suboptimal results.
The new generation of Claude models, based on my observations, has learned to independently interpret the surrounding context. That is why Anthropic engineers concluded that rigid, deterministic rules are no longer needed. Reducing the prompt by 80% is not just optimization but an acknowledgment that the model has become sufficiently "mature" to make decisions on its own.
New rules for working with Claude
Instead of multi-page instructions, Anthropic proposes four simple principles:
1. Trust the model's judgment. Instead of prohibitions (e.g., "don't write comments in code"), a single line suffices: "write code in the style of the surrounding project."
2. Limit the number of examples. Previously, few-shot examples were the gold standard. Now, on the contrary, they constrain the model, forcing it to blindly copy the shown pattern instead of seeking the optimal solution.
3. Don't overload the context. There is no need to cram all instructions into one file. Claude will pull up the necessary information at the right moment on its own. Extra rules only consume tokens and create noise.
4. Say it once. Previously, instructions were duplicated in different places—in the prompt, in tool descriptions. Now duplication is considered harmful. The model will find and apply the needed rule on its own.
Practical takeaway
My analysis shows that we are witnessing a tectonic shift in the paradigm. From "prompt engineering" (the art of crafting the perfect query), we are moving to "context engineering"—the art of creating the right environment in which the model can unlock its potential. The system prompt is only important when building your own agents; for the average user of Claude Code, it is no longer a critical element.
This is a signal for everyone working with AI: fewer instructions, more freedom. The new generation model can separate the wheat from the chaff on its own. Our task is not to hinder it but to provide high-quality yet minimally necessary context. This is not just a trend; it is a new reality that will change the approach to AI-driven development. The market, which tracks the effectiveness of AI tools, should take note of this.