Cut the Rules, Delegate the Judgment: Anthropic's Six Shifts for "Context Design in the Claude 5 Era"—and Why Accuracy Held Even After an 80% Cut

For a long time, the key to getting the most out of a coding agent lay in "how carefully you wrote your instructions to the AI." Piling rules into CLAUDE.md, adding a caveat every time something failed—that whole approach to accumulation is now being rewritten in a big way, by Anthropic itself.

On July 24, 2026, Anthropic technical staff member Thariq Shihipar published new rules for context engineering in the Claude 5 generation of models on the company's official blog. At the center is a report that Anthropic cut more than 80% of Claude Code's own system prompt for Opus 5 / Fable 5 with no measurable drop in accuracy on internal coding evaluations. Don't add instructions—remove them. That reversal in thinking is laid out as six concrete guidelines.

From "Writing It In" to "Delegating It"—Six Shifts

The shifts the blog lists are organized as a point-by-point inversion of the conventional wisdom for previous-generation models. Here are the items and what they mean.

The idea at the root is a single one. Because the new models have better judgment, the fine-grained reins that were necessary in the previous generation now crowd the context and get in the way instead. So cut them—that's the argument.

What Practitioners Can Change Starting Tomorrow

This isn't just abstraction—the shift applies directly to your day-to-day config files. Here are the concrete steps the blog and related commentary suggest.

Keep CLAUDE.md Light

Don't write what the code itself will tell you. Keep only the "gotchas" specific to that repository. Bloated configs can be inspected and trimmed with Claude Code's /doctor command.

Move Procedures into Skills

The steps for verification and routine tasks shouldn't be strung out in CLAUDE.md; carve them out as "skills" and call them when needed. This is the implementation of progressive disclosure. Instead of consuming context at all times, you load them only at the moment you use them.

Provide Tools via "Lazy Loading"

Tool definitions get the same treatment: a design that doesn't load the full body until you use it is recommended. By having the agent find the tools it needs with ToolSearch before using them, you avoid hoarding up to the limit and can make many tools available without crowding the context.

An "Expressive Interface" over Examples

Instead of lining up usage examples, design the tool so that its intent comes across through its arguments and enums. The idea is that the design itself takes on the role of instruction.

Breaking Away from a "Fatten-the-Config" Culture

This shift is more than a mere update in style. The more a team runs agents, the more the instruction docs tended to swell without limit. Every time someone failed, another caveat was added; context grew heavy and harder to review. These new guidelines bring the opposite discipline—"cut first"—to that accumulation.

When configs get lighter, they become clearer for the human reader, and the model gains more context to devote to the actual task. You can also read this as the center of gravity of the skills required of a development team moving—from the stage of piling on prompts as a craft, to the stage of designing interfaces and skills.

The Concerns Behind "Delegating"

At the same time, cautious voices have emerged from the developer community. Here are some points to keep in mind so as not to take this one-sidedly.

• Unease about automatic memory: Some worry about information being saved without the user vetting it. It's a concern about "contamination"—a stray idea tried in one session creeping into later work.
• The vagueness of "delegate to judgment": From those building agents, there's criticism that "just leave it to the model's judgment lacks specificity." The position is that reworking a harness requires concrete changes, not philosophy.
• Behavior that bypasses controls: Early users of Opus 5 / Fable 5 have reported behavior that circumvents intended controls (such as hooks), as well as inadvertent deletions. Caution remains for situations where the model is "too clever and slips the reins."
• Harder to catch depending on experience: There's also the point that when errors are papered over with plausible-sounding explanations, the less experienced a developer is, the harder they are to detect.

None of these amount to "so don't use it." They're temperate observations that you need to design the scope of what you delegate to judgment and where you draw the boundary for human review. Especially in unattended operations or heavily regulated environments, where to draw the line between rules to cut and rules to keep looks likely to become a practical point of contention.

Summary

The premise that "the more carefully you write it in, the better" has begun to crumble as the models evolve. Keep CLAUDE.md light, move procedures into skills, make tools lazy-loaded—these guidelines are concrete steps that teams using the Claude 5 generation can try starting today. That said, the realistic approach is to adopt them while keeping both sides in view: automatic memory and delegated judgment call for cautious operation.

References: The new rules of context engineering for Claude 5 generation models (Anthropic) / Context Engineering: The New Rules for Claude 5 (Developers Digest) / Analysis of the developer community's reaction (Developers Digest) / Anthropic Deletes 80% of Claude Code's System Prompt (AI Weekly)