In the world of AI development, a telling paradigm shift has occurred. Anthropic, after analyzing the internal telemetry of its Claude Code tool, came to an unexpected conclusion: more than 80% of the system prompt, which for years was considered critical for controlling the model, turned out to be not just useless, but harmful. This is about a fundamental revision of the principles of context engineering for the new generation of Claude models.

Why the old rules worked against us

The problem lay in contradictions. The traditional approach involved saturating the model with detailed instructions, skills, and examples. However, log analysis showed that these components conflicted with each other. The system prompt demanded one thing, the tool description demanded the opposite, and the user request added a third, incompatible vector. The model, overloaded with contradictory directives, spent computational resources trying to reconcile them rather than solving the task.

Anthropic's key insight is that the new Claude models have become "smart" enough to independently interpret the surrounding context. They no longer need rigid frameworks and multi-page manifestos. Instead, they are capable of making decisions based on minimal but clear signals. That is why the company went for a radical reduction, and notably, in internal programming tests, the quality of results not only did not drop, but in a number of scenarios even increased.

New rules: less text, more trust

Anthropic's new philosophy is built on four principles that overturn previous notions of prompt engineering. First, the model needs to be trusted. Prohibitions like "don't write comments in the code" are replaced with a simple recommendation: "follow the style of the surrounding code." Second, examples are no longer the gold standard. Three demonstrated ways of solving a problem make the model perceive them as the only possible ones, which limits its creativity and adaptability.

Third, instructions do not need to be dumped into one giant file. Claude will pull up the necessary information on its own at the right moment, while extra rules only burn tokens and create noise. Fourth, duplication is evil. If the same rule is mentioned in the prompt, the tool description, and memory, it creates a conflict of priorities. Now Anthropic recommends removing all repetitions, leaving only the single clearest version.

Special attention deserves the advice on working with memory and artifacts. Instead of simple markdown files, which the model may interpret ambiguously, it is better to use HTML and structured data. This allows the model to extract the needed information faster and more accurately.

My analysis: This step by Anthropic is not just optimization, but a signal of the maturation of LLMs. We are witnessing a transition from the era of "training" models with rigid prompts to the era of "mentoring," where trust in the model's reasoning plays a key role. For developers, this means it is time to revisit their CLAUDE.md and system prompts, leaving only what is most important. Otherwise, you are not just wasting tokens, but actively preventing the model from working at full capacity. This is a lesson worth learning for everyone building agentic systems on modern AI.