Anthropic’s post on context engineering for Claude 5 generation models leads with a number that does most of the arguing: over 80% of Claude Code’s system prompt was removed for Opus 5 and Fable 5 class models with no measurable loss on evals.
The six shifts it describes compress into one sentence: stop scripting the model’s behavior and start curating what it sees. Explicit rules give way to model judgment, upfront context dumps to progressive disclosure, repeated instructions to clean tool descriptions, manual CLAUDE.md updates to automatic memory, and prose specs to rich references - an HTML mockup beats a description of one.
Two things stand out for anyone running a file-based agent workflow.
First, the guidance codifies what file-based setups already practiced: lightweight per-repo CLAUDE.md focused on repository gotchas, skills split into small files loaded on demand. If your context lives in versioned markdown routed in at the right time, you have been doing progressive disclosure all along - the argument I made in AI skills as workflow infrastructure.
Second, the memory shift is the consequential one. Retiring the manual hotkey in favor of automatic memory creates a third category between hand-written files and opaque stores: file-based auto-memory, readable and scoped per repository but stored outside it, so still not reviewable in a PR. That is the custody line from Why git context beats memory agents, and this release moves the industry closer to it without crossing it.
The anti-pattern list alone is worth the read: overconstrained guidance, duplicated instructions, context loaded upfront, and CLAUDE.md treated as a central knowledge repository. The new /doctor command in Claude Code now trims exactly these.
Original post: The new rules of context engineering for Claude 5 generation models.
Source: claude.com