Replace markdown specs with executable references — code, test suites, and rubrics beat prose descriptions for Claude 5 models
Claude by Anthropic·medium signal
The same guidance says to stop front-loading all information and instead use progressive disclosure across multiple skill files, loading context only when needed. It explicitly downgrades "simple markdown specs" in favor of rich references: HTML artifacts, code, test suites, and scoring rubrics, with @mentions to pull in detail on demand. CLAUDE.md is repositioned as a lightweight file that should carry codebase gotchas rather than restate obvious facts, and skills are where team-specific opinions belong.