AI Insights · Oct 22, 2026 · Verified

Two small files now carry agent instructions everywhere

SKILL.md and AGENTS.md now work across vendors: Codex reads the same skills standard, and Claude Code has read AGENTS.md since September 2026.

Claim

Two tiny formats now carry agent instructions across vendors. SKILL.md (name plus description required, about 100 tokens loaded at startup, body under 5k tokens on activation) is read by Claude, Codex, Cursor, OpenCode, Amp, Letta, goose, GitHub and VS Code. AGENTS.md (nearest file wins, 60k-plus projects) is read by 20-plus tools and by Claude Code when no CLAUDE.md exists. CLAUDE.md and .cursor/rules remain the vendor-primary files.

Evidence

  • Agent Skills spec: name (64 chars, matches directory), description (1024 chars), optional license, compatibility, metadata, experimental allowed-tools; scripts/, references/, assets/; three loading tiers. Willison called it "deliciously tiny" and under-specified on metadata.
  • Codex loads skills from .agents/skills (repo, parents, home, /etc/codex/skills) and says it builds on the agentskills.io standard; optional agents/openai.yaml adds UI metadata and policy. Directory divergence: Anthropic .claude/skills, OpenAI .agents/skills.
  • AGENTS.md was contributed to AAIF by OpenAI; OpenAI's main repo has 88 of them; Codex caps combined AGENTS.md at 32 KiB.
  • Claude Code 2.1.277 (2026-09-18) reads AGENTS.md when no CLAUDE.md exists, implemented as a built-in mod. A CLAUDE.local.md silently disables the fallback (secondary). UNVERIFIED.
  • Stripe found modular progressive-disclosure skills beat one monolithic file by about 10% and cut tokens (see harness-engineering findings).

Open questions

  • Whether Agent Skills will be donated to AAIF; no evidence yet. Answered 2026-10-23: Anthropic donated the standard to AAIF on 2026-10-22, see [[2026-10-agent-skills-donated-to-aaif]]. Still open: the first versioned release and the security guidance the maintainers proposed.

Sources

AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.

Book a demo