The new rules of context engineering for Claude 5

The Claude team says the arrival of the Claude 5 generation, Opus 5 and Fable 5, changed how they write context for Claude Code. The headline example: they removed more than 80 percent of Claude Code's system prompt and saw no measurable drop on their coding evaluations. The lesson they draw is that stronger models need fewer rigid rules and can be trusted to use judgment, so much of the old scaffolding was just noise.

The post lays out five shifts. Prefer judgment over hard rules: instead of an absolute "DO NOT add comments," tell the model to match the surrounding style. Design expressive tool interfaces rather than piling in usage examples that box the model in. Load context only when it is needed, through skills, deferred tool loading, and tiered files, rather than front-loading everything. Cut repetition, because the same instruction repeated across the system prompt and tool descriptions used to create conflicts, like "leave documentation as appropriate" fighting "DO NOT add comments." And lean on rich references like code, tests, and rubrics instead of plain markdown specs.

There are concrete product changes too. Memory now auto-saves instead of asking you to hand-edit CLAUDE.md, and a new /doctor command helps trim skills and context files that have grown too large. The team's advice is to keep CLAUDE.md light and reserve it for repository-specific gotchas, not things the model already knows.

Why it matters

If you maintain long system prompts or bloated CLAUDE.md files for coding agents, much of that effort may now be hurting more than helping. Try deleting rules and examples on Claude 5 models and measure the result, rather than assuming more instruction is always safer.