Tip: Keep AGENTS.md Under 150 Lines as an Index, Push Situational Knowledge Into `.agents/skills/`
Red Hat published concrete sizing guidance on 2026-07-27 for splitting always-on context from situational context: AGENTS.md is sent with every prompt so it should stay under 150 lines (30–50 for small repos) and act as an index — "detailed architecture is at docs/ARCHITECTURE.md" — rather than a dump, with the inclusion test being "would removing this cause the agent to make a mistake?" Everything episodic moves into SKILL.md files whose name and description alone load into context, with full body read only on trigger. The portability tip that matters: store local skills in `.agents/skills/` rather than a vendor path for maximum cross-agent compatibility, and have CLAUDE.md reference `@AGENTS.md` plus Claude-specific additions instead of duplicating it. Skill instructions should be imperative ("Run linting") not tool-prescriptive ("Use the Bash tool to...").
↳ Follow the thread