Anthropic rewrites context engineering for Claude 5: Less rules, more judgment

Anthropic has fundamentally rewritten the playbook for how developers should guide its latest generation of AI models. In a detailed post by Anthropic staff member Thariq Shihipar, the company reveals it removed over 80% of Claude Code's system prompt for its newest models, including Claude Opus 5 and Claude Fable 5, with no measurable loss in coding evaluation performance.

The shift represents a major philosophical change in how to approach what Anthropic calls "context engineering"—the assembly of system prompts, skills, CLAUDE.md files, and memory that shape a model's behavior. Instead of piling on constraints, the new mantra is to trust the model's judgment.

The end of over-constraining

Anthropic found that its previous approach was actively harming performance. Internal transcripts of Claude Code usage showed conflicting messages in single requests, such as "leave documentation as appropriate" alongside "DO NOT add comments," forcing the model to spend cognitive effort reconciling contradictions.

"We were overconstraining Claude Code, both through our system prompt and in our CLAUDE.md files and skills," Shihipar wrote. The new generation of models, which includes the recently released Claude Opus 5 and the earlier Claude Fable 5, now has better judgment and can handle decisions without explicit rules.

On benchmarks, the improvement is stark. Claude Opus 5 scored 43.3% on FrontierBench v0.1 at max effort, compared to 18.7% for Opus 4.8 and 33.7% for Fable 5. At the highest 'xhigh' effort setting, Opus 5 reached 44.4% mean reward.

From rules to interfaces

One of the most significant changes involves how developers should design tools for Claude. The old rule was to provide examples of how to use tools. The new rule: design better interfaces.

"With our newest models, we've found that giving examples actually constrains them to a certain exploration space," Shihipar explained. Instead, developers should think about the design of their tools, scripts, and files—what parameters does Claude have, and how can they be more expressive?

For instance, a simple enumeration of status values like pending, in_progress, and completed hints to Claude how to use a tool without needing explicit instructions. The company also recommends putting tool instructions in the tool descriptions themselves rather than in the system prompt.

Progressive disclosure and auto-memory

Anthropic now advocates for progressive disclosure—loading the right context at the right times rather than front-loading everything. Claude Code has become competent at selectively calling skills and tools only when needed, using features like deferred loading and ToolSearch.

"A common myth is that you want to make your CLAUDE.md a central repository for every known practice that you might run into," Shihipar noted. Instead, developers should consider a tree of files that can be loaded at the right time.

For memory, the old practice of manually saving notes to CLAUDE.md files has been replaced by auto-memory. Claude now automatically saves relevant memories without user intervention, eliminating the need for manual note-taking through the # hotkey.

Practical advice for developers

Anthropic's guidance for developers using Claude Code or building their own agents includes several key changes:

• Keep CLAUDE.md lightweight: Focus on codebase-specific gotchas, not obvious information. Avoid stating what Claude can see in the file system.
• Design skills as guides: Skills should encode particular opinions or best practices, but avoid over-constraining them. Use progressive disclosure for long skills by splitting them into multiple files.
• Use rich references: Instead of simple markdown files, Claude can now reference HTML artifacts, code, test suites, or even entire codebases. A spec might be a detailed test suite rather than a document.
• Run `claude doctor`: Anthropic has released a new command that automatically trims bloated setups, removing outdated constraints and contradictions.

For those building their own agent harnesses, the system prompt remains critical. "A system prompt is heavily tied to the product context. It tells Claude what product it's operating in and what it's doing," Shihipar wrote.

Breaking changes and model specifics

The release of Claude Opus 5, which supports a 1 million token context window and 128k token maximum output, brings notable changes. Thinking is now on by default, and the effort parameter controls depth. A breaking change: setting thinking: {"type": "disabled"} with effort xhigh or max returns a 400 error.

Anthropic also warns developers to delete their verification prompts. Instructions like "include a final verification step" now cause over-verification, because the model already verifies its own work. The minimum cacheable prompt drops to 512 tokens, down from 1,024, making caching more accessible.

One notable benchmark detail: Opus 5 safety classifiers flagged and refused 5% of API calls across 4% of trials, while Fable 5 classifiers flagged 42% of calls across 26% of trials—a significant difference in safety behavior between the two models.

The bottom line

The overarching message from Anthropic is clear: stop treating the model like a rule-following machine and start treating it like a reasoning agent. "Give it the goal, not a rulebook—let it use judgment," Shihipar advises. "When it over-does things, remove instructions—don't add them."

For developers who have accumulated years of prompting wisdom for older, weaker models, the advice is simple: clear out stale rules, delete contradictions, and re-test. The new generation of Claude models doesn't need the training wheels anymore.