July 25, 2026, (Inside AI) — Anthropic has overhauled its context engineering playbook for the new Claude 5 generation models, discarding rigid rules in favor of design that trusts the model's judgment. In a technical blog post, Thariq Shihipar, a member of technical staff at Anthropic, detailed how practices once essential for safety and performance are now myths that can hinder the advanced reasoning of Claude 5.
The shift is most dramatic in system prompts. Where older models needed explicit, sometimes overbroad guardrails—like a directive to “default to writing no comments” in code—Claude 5 can follow nuanced instructions. The old prompt warned: “Never write multi-paragraph docstrings or multi-line comment blocks -- one short line max.” But this would fail for complex code or user preferences. Now, the system prompt simply says: “Write code that reads like the surrounding code: match its comment density, naming, and idiom.”
This change reflects a core principle: earlier models required rules to avoid worst-case scenarios, such as deleting files. Shihipar explained that for Claude Code, “we needed to be sure that Claude avoided worst case scenarios,” so they accepted the tradeoff of occasionally incorrect guidance. But newer models “have better judgement and can handle these decisions well without explicit rules.”
Tool usage has also been reimagined. Previously, providing examples was the top rule for teaching Claude how to use tools. Now, Anthropic found that “giving examples actually constrains them to a certain exploration space.” Instead, the focus is on designing expressive interfaces. For instance, a Todo tool with a status parameter that enumerates “pending, in_progress, and completed” hints at its use, while an instruction to keep one item in_progress defines desired behavior without explicit examples.
Progressive disclosure replaces the old practice of front- all information. In Claude Code, detailed code review instructions were once crammed into the system prompt. Now, those are moved into separate skills that the model can call selectively. Shihipar noted that “progressive disclosure is not just for skills, we also use it for tools,” with some tools using deferred so their full definitions are fetched only when needed. This keeps the context window lean.
The same applies to user-defined CLAUDE.md and Skill.md files. Rather than a monolithic repository of every practice, Anthropic recommends a tree of files loaded at the right time. Memory management also evolves: instead of users manually saving to CLAUDE.md via a hotkey, Claude now “automatically saves memories that are relevant to the work and to you.”
For planning, simple markdown files give way to rich references. Claude can now use HTML artifacts, detailed test suites, or even functions from other codebases as specs. Rubrics—dynamic workflows that spin up verifier agents—let Claude assess quality against a user’s taste, such as what good API design looks like. Shihipar’s post culminates in a practical guide for assembling context. System prompts should be product-specific and rarely modified in Claude Code. CLAUDE.md files should be lightweight, focusing on codebase “gotchas” rather than the obvious, and lean heavily on progressive disclosure. Skills should be lightweight guides that encode team-specific knowledge, avoiding overconstraint except in critical areas. References—@-mentioned files—should be in code when possible, as “a HTML mockup of a design will generally produce better results than a description of the design or a screenshot.”
To help users adapt, Anthropic rolled out a new command, “claude doctor,” which automatically simplifies context files. The company also pointed to its Fable field guide for more prompting details. These updates align with broader research on in-context learning and model scaling, where larger models require less explicit instruction and benefit from structured, retrievable context—a trend documented in studies like “Larger language models do in-context learning differently”.
Why Guardrails Became a Liability #
The old system prompt for Claude Code included blunt rules that sometimes backfired. Shihipar recalled the directive to avoid comments, which was meant to prevent verbose, incorrect documentation. But for complex code or documentation tasks, it was counterproductive. The new approach trusts Claude 5 to infer intent from surrounding code, a capability that earlier models lacked. This mirrors a broader industry shift from hard-coded rules to adaptive, context-aware behavior in large language models.
Tool design follows the same philosophy. By removing examples, Claude 5 explores a wider solution space. The Todo tool’s design is a case study: the enumeration of states and the “one item in_progress” rule communicate intended workflow without step-by-step examples. This reduces prompt size and avoids anchoring the model to specific patterns.
Progressive disclosure is perhaps the most impactful change. In earlier versions, the system prompt was a dense manual. Now, skills and tools are loaded on demand, akin to how an operating system manages memory. Shihipar emphasized that this allows more tools without bloating the context, a critical advantage as agentic systems grow in complexity.
From Static Memory to Adaptive Context #
Auto-memory marks a departure from user-managed context. Previously, users had to explicitly save information to CLAUDE.md; now, Claude identifies and stores relevant details autonomously. This reduces cognitive load but raises questions about transparency—Anthropic has not detailed how these memories are prioritized or reviewed.
The move to rich references also signals a maturation of AI-assisted development. By using code-based specs and HTML mockups, Claude can leverage its training on structured data for higher fidelity outputs. Rubrics add a layer of automated quality control, enabling verifier agents to check work against predefined standards. This could accelerate complex projects but requires careful rubric design to avoid overfitting.
Shihipar’s advice to simplify context files is a final nod to the new paradigm. The “claude doctor” command automates this pruning, but the underlying message is clear: with Claude 5, less is more. As models improve, the burden shifts from prompt engineering to interface design—a trend that will likely define the next generation of AI tools.