Anthropic has published new guidance telling developers to strip rules out of their Claude prompts rather than add more. The company says its newest models handle ambiguous situations well enough that the heavy-handed constraints written for earlier Claude versions now do more harm than good. This is vendor guidance from the vendor whose models it describes, and it doubles as an admission that a lot of prompt engineering built for older Claude releases is now dead weight.
The specific example Anthropic gives involves code comments. Earlier system prompts for Claude Code told the model to default to zero comments and never write multi-line docstrings, a blanket rule needed because older models could not tell when documentation was actually warranted. Anthropic says newer models can make that call themselves, so the current instruction simply asks Claude to match the style of surrounding code.
Anthropic is also walking back its own advice on tool examples. The company used to recommend showing Claude sample usage for every tool. It now says examples narrow the model’s exploration space more than they help, and recommends designing clearer tool interfaces instead, such as an enum field with values like pending, in_progress, and completed that signals intended behavior without a worked example attached.
The bigger structural change is what Anthropic calls progressive disclosure: loading instructions only when a task actually needs them, instead of front-loading a system prompt with every rule that might apply. Claude Code now keeps verification and code review logic in separate skills the model calls on demand, and some tools use deferred loading, meaning Claude has to search for a tool’s full definition before it can use it. Anthropic also says repeated instructions, once used because older models paid more attention near the end of a context window, are no longer necessary, and that Claude now saves relevant memories automatically rather than relying on users to write them into a CLAUDE.md file by hand.
None of this is neutral advice. Anthropic is describing tradeoffs it made while building its own product, and the guidance conveniently favors the exact context-management features (skills, deferred tool loading, auto-memory) that Anthropic has shipped into Claude Code over the past year. The blog post does not include independent testing showing rule-based prompts actually perform worse on other harnesses, only Anthropic’s account of its own internal rollout.
For teams that already built prompt libraries around the old model generation, this is a rework project, not a footnote. Any CLAUDE.md file padded with defensive rules, any tool description duplicated across a system prompt and its schema, any workflow leaning on repeated instructions to fight context-window recency bias needs an audit against models that no longer need that scaffolding. Anthropic’s own fix, a new claude doctor command meant to help simplify bloated prompt files, suggests the company expects this cleanup to be common enough to automate. Teams running dozens of agent configurations should expect this to consume real prompt-engineering time, likely spread across several sprints rather than a single afternoon, before their setups actually benefit from the judgment newer Claude models are supposed to have.
Any team maintaining Claude-based agents should schedule a prompt audit before the next model upgrade lands, since rules written for judgment-poor models may now be actively suppressing the capability they paid for.
This account is based on Anthropic’s blog post “The new rules of context engineering for Claude 5 generation models,” published on the Claude blog.
