The new rules of context engineering for Claude 5 generation models
Claude ● Covered by 17 sources
Anthropic just rewrote its own playbook for how to prompt Claude Code, because the old tricks stopped working on newer models. Turns out spelling everything out and repeating instructions now hurts more than it helps.
Based on reporting by Claude — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Anthropic has quietly rebuilt the rulebook for how it talks to its own coding agent, and the changes say a lot about how fast these models are moving. Thariq Shihipar, a member of technical staff at the company, laid out six practices that used to be gospel for prompting Claude Code and are now, in his words, myths.
The biggest shift is trust. Early versions of Claude needed rigid rules — things like banning multi-line comments outright — because the model couldn't be trusted to make judgment calls without breaking something. Newer models can. So Anthropic ripped out the hardcoded restrictions and replaced them with softer guidance, like telling Claude to match the comment style of whatever code already surrounds it. Same story with tool examples: instead of spoon-feeding sample usage, which Shihipar says actually narrows how creatively the model explores a problem, the team now spends its effort designing better tool interfaces and lets Claude figure out the rest.
Context management got an overhaul too. Anthropic used to cram everything into the system prompt up front, on the theory that Claude might need any of it at any moment. Now it leans on progressive disclosure — loading information only when it's actually relevant. Verification and code review, once baked into every prompt, now live in separate skills that Claude calls up on demand. Some tools use a feature called deferred loading, meaning their full definitions only surface through a ToolSearch lookup, which frees up context space for tools that would otherwise sit there unused. The same logic applies to CLAUDE.md files: rather than one giant document trying to anticipate every scenario, Anthropic recommends a tree of smaller files pulled in only when needed.
A few other habits got dropped for similar reasons. Repeating instructions in both the system prompt and the tool description used to compensate for models that paid more attention to the end of their context window — that redundancy is gone now that tool descriptions alone do the job. Manual memory-saving via a hotkey has been replaced by Claude automatically deciding what's worth remembering. And simple markdown spec files are giving way to richer references: HTML artifacts, test suites, even rubrics that let Claude spin up verifier agents to check whether something matches a particular taste in, say, API design.
Shihipar's practical advice for anyone building on top of this: keep CLAUDE.md files short and focused on genuine gotchas rather than obvious facts about the codebase, treat skills as narrow expert guides rather than sprawling manuals, and prefer code-based references over prose or screenshots because Claude simply reads code better. Anthropic even shipped a new command, claude doctor, to help developers simplify their own bloated prompt setups the same way the internal team did.
My take — AI-written commentary, not fact-checked reporting
This is basically Anthropic admitting its early prompting advice was a workaround for a model that wasn't smart enough yet, and I appreciate the candor. The pattern here — rigid rules giving way to judgment, upfront context giving way to on-demand retrieval — is the same trajectory every maturing AI product goes through, and it's a good reminder that most of today's prompt-engineering folklore has a shelf life measured in model generations, not years.
Read more about this at: Claude