Anthropic says it has sharply reduced the amount of standing instruction supplied to advanced Claude Code models, presenting the change as evidence that context engineering must evolve alongside model capability. The company reports removing more than 80% of the system prompt used with Claude Opus 5 and Claude Fable 5 without measuring a loss on its coding evaluations.
The change focuses attention on the difference between a user prompt and the broader context assembled around it. In Claude Code, that context can include a system prompt, skills, `CLAUDE.md` files, memory and other resources. Because these materials may apply across many different requests, overly specific rules can conflict with one another or constrain a model in situations their authors did not anticipate.
Anthropic says its internal transcripts showed overlapping instructions, including general requests to leave suitable documentation alongside rigid directions not to add comments. Newer models could often infer the intended result, but resolving those clashes required additional reasoning. Rules originally intended to prevent poor outcomes with earlier systems had therefore become unnecessary or counterproductive as model judgment improved.
One example is coding style. An older instruction discouraged comments, multi-paragraph docstrings and unsolicited planning documents. Anthropic replaced that prescriptive language with a shorter direction to match the surrounding code’s comment density, naming and idiom. The newer formulation leaves the model room to respond to a complex codebase or a user’s explicit preferences rather than applying a universal prohibition.
The company also revised its view of examples. Detailed tool-use demonstrations once helped models understand expected behavior, but Anthropic says examples can now narrow exploration. It recommends expressive tool definitions and parameters that communicate constraints directly. An enumeration of task states, for instance, can indicate permitted values without a lengthy sample dialogue.
Progressive disclosure is another part of the approach. Verification and code-review guidance can live in skills loaded when relevant, while deferred tools need not occupy context before use. Anthropic similarly advises against treating one project instruction file as a store for every possible practice. A tree of focused files can make specialized material available on demand.
The company has put these recommendations into Claude Code’s `/doctor` command, which it says can help users resize skills and instruction files. The broader message is not that context no longer matters. Instead, Anthropic argues that useful context should be concise, nonconflicting and accessible at the moment it is needed, with fewer blanket rules standing between a capable model and the specific task. Users still need firm safeguards where consequences demand them; the recommendation is to make each instruction earn its place and avoid stale duplication.
