Claude models have become even more advanced, and your context engineering rules needed to be revolutionized. Recently, the development team noticed a huge leap in how the latest models, such as Claude Opus 5 and Claude Fable 5, are triggered. Over 80% of the Claude Code system prompt was removed without any measurable loss in coding evaluations.
Context engineering in Claude involves creating general prompts that are used across multiple requests and cannot be as specific as a common prompt. With the evolution of Claude’s own capabilities, it has become challenging to build these general prompts. Claude can interpret the user’s intent to arrive at the correct answer, but it must think more carefully about conflicting and overlapping messages before deciding what to do.
In this context, it became possible to remove many of these constraints and allow the model to use surrounding context and judgment instead. Claude Code now has more tools. Previously, Claude relied on the CLAUDE.md file as a source of memory, information, and guidance. Now we have memory, artifacts, and skills that Claude can use to create new ways to load and share context across sessions.
Past best practices of context engineering have turned into myths, such as the need to give Claude rigid rules, give examples, or put all necessary information upfront. Today, we advise letting Claude use judgment, designing interfaces, using progressive disclosure, and describing tools simply.
To apply this to your context, consider the following: a system prompt is strongly tied to the product context and tells Claude which product it is operating in and what it is doing. As for CLAUDE.md, keep it light and briefly describe what your repository is for, but use most tokens for peculiarities within the codebase.
These changes in Claude Code pave the way for smarter and more flexible context engineering, allowing Claude 5 models to work with efficiency and reliability never seen before.