In the rapidly developing world of artificial intelligence, a landmark event has occurred. Anthropic, a company known for its advanced Claude models, has taken a seemingly paradoxical step. During a deep analysis and optimization of its own programming tool, Claude Code, more than 80% of the system prompt was removed. This decision is not just a technical detail, but a fundamental shift in the philosophy of interacting with large language models.
Why Less Became More
My analysis shows that the reason lies in so-called context engineering. Previously, it was believed that the more instructions and rules we load into a model, the more accurate its work would be. However, Anthropic engineers, studying internal logs, discovered a critical problem: their own instructions directly contradicted each other. The system prompt, the skill set, and the user request conflicted, creating "cognitive dissonance" in the model.
One part of the context demanded one thing, another demanded the exact opposite. This led to instability and a decline in code quality. As it turned out, the new generation of Claude models is capable of independently analyzing the surrounding context and making decisions. That is why it was possible to abandon a huge array of outdated rules without losing performance. In internal programming tests, quality not only did not drop but remained at a high level.
New Rules of the Game
Instead of rigid and verbose instructions, Anthropic offers a new, more concise approach. It is built on four key principles that I consider revolutionary:
1. Trust the model's judgment. Instead of prohibitions (for example, "don't write comments") — one simple recommendation: "write code in the style of the surrounding project."
2. Examples are the enemy of freedom. Previously, demonstrating three examples was considered the gold standard. Now, they limit the model, forcing it to copy the proposed templates instead of finding the optimal solution.
3. Gradual information disclosure. There is no need to dump all instructions into one giant file. The model will pull in the necessary information at the right moment itself. Extra rules only waste tokens and create noise.
4. Say it once. Outdated models required repetitions to absorb information. Now, duplicating instructions is harmful. If a rule exists in one place, it should not exist in another.
Practical Takeaways for Developers
My main advice, based on this experience, is to audit your settings. Ask yourself three questions: can the model handle this without this instruction? Is it duplicated somewhere else? Is it needed always or only in rare cases? If the answers are obvious, it is better to remove the rule or move it to a separate skill that will only be loaded when necessary.
This step by Anthropic is an important signal for the entire industry. We are moving from the era of "prompt engineering" to the era of "context engineering." Less is more — and this is not just a slogan, but an effective strategy that allows us to unlock the true potential of modern neural networks. Free the model from the shackles of unnecessary rules, and it will show what it is truly capable of.