Tobi said Claude Code insisting on CLAUDE.md causes split-brain when teammates use other tools that read AGENTS.md and .agents/skills. The docs agree: Claude Code reads CLAUDE.md, not AGENTS.md. The recommended fix is a CLAUDE.md that contains @AGENTS.md, or a symlink. bcherny closed #6235 as completed on that workaround on 2026-08-17.

That import loads the instruction file at session start. It does not discover skills. Claude Code’s skills docs still list the default user and project paths as ~/.claude/skills/ and .claude/skills/. The Agent Skills client guide treats .agents/skills/ as the cross-client scan path; the spec does not require it. /import appends a one-time copy of AGENTS.md into CLAUDE.md. /init with CLAUDE_CODE_NEW_INIT=1 can read AGENTS.md when it generates a CLAUDE.md. Neither is dual-discovery. An AGENTS.md-only repo still needs the extra file.

I would score instruction_file, skills_root, and loaded separately. I did not run Claude Code against an AGENTS.md-only tree.