Anthropic’s new advice for Claude: stop telling it everything

Anthropic says the latest generation of Claude models performs better with shorter, clearer background instructions and more room to apply its own judgement. The company reports that it removed more than 80% of the Claude Code system prompt for models including Claude Opus 5 and Claude Fable 5, without a measurable decline in its coding evaluations.

Thariq Shihipar writes for the official Claude Blog that the company found its own guidance had become overly restrictive and sometimes contradictory. The findings matter beyond software development because many AI tools combine a user request with persistent instructions, saved preferences, reference files and task-specific guidance.

Anthropic calls this wider collection of information “context engineering.” It includes the system prompt, which defines the product’s role, as well as project files, memories, skills and attached materials. Unlike a single prompt, this context must work across many possible tasks. That makes excessive detail particularly risky.

According to Anthropic, older models needed firm rules to avoid failures such as deleting files or writing unsuitable comments in code. Those safeguards could also force bad decisions in situations where the rule did not fit. A blanket instruction never to write detailed comments, for example, may conflict with a documentation task or a complex part of a project that needs explanation.

The company now recommends providing a broader principle rather than a long list of prohibitions. In a code project, that could mean asking Claude to match the surrounding style, including its naming patterns and amount of commentary. The model can then use the local context to decide what is appropriate.

From detailed instructions to better context

Anthropic identifies several changes in how teams should structure AI guidance:

• Use judgement instead of rigid rules: Remove generic instructions that repeat obvious expectations or conflict with user requests.
• Design clear interfaces instead of supplying examples: For tools and workflows, expressive options and clearly named states can guide the model without narrowing its choices to a few sample cases.
• Load information when needed: Keep specialised material in separate files or skills, rather than placing every instruction in one large permanent document.
• Put tool guidance with the tool: Anthropic says teams can often remove repeated reminders from the overall system prompt when the relevant tool description already explains correct use.

This approach, often called progressive disclosure, is central to the company’s advice. A model does not need detailed review rules, verification steps or specialist information for every request. Instead, it should be able to retrieve those materials when the task calls for them. Anthropic says Claude Code uses separate skills for activities such as code review and verification, and some tools load their full definitions only after the model searches for them.

For organisations building their own Claude workflows, Anthropic recommends keeping general project instruction files lightweight. Such files should briefly explain the purpose of a repository or workspace and focus on exceptions that the model cannot easily infer, such as unusual file structures, internal conventions or required checks.

The company also advises teams to treat skills as focused guides rather than exhaustive manuals. Long guidance can be split into linked files. The most valuable material is information that reflects a team’s particular practices, product knowledge or quality standards.

Anthropic further argues that rich reference materials can outperform simple written specifications. An AI model may work more effectively from a detailed test suite, existing code, a design mock-up in HTML or a rubric that defines quality criteria. These assets show the desired outcome in a form the model can inspect directly.

The practical message is straightforward: do not assume that more instructions produce better AI work. For capable models, carefully selected context, clear tools and relevant references may be more useful than a large rulebook.