OpenAI's playbook for making Codex less forgetful

What it does

This repository catalogs reusable “skills” — folders of instructions, scripts, and resources — that OpenAI’s Codex agent can discover and apply to specific tasks. The pitch is “write once, use everywhere”: package a capability, share it across a team, and stop re-explaining the same workflow to your agent.

The interesting bit

The structure is deliberately tiered. .system skills ship automatically with Codex, .curated ones install by name, and .experimental ones live in the wild-west folder for trying things out. There’s also an open standard at agentskills.io, suggesting OpenAI wants this format to outlive Codex itself.

Key highlights

• Skills are plain folders with instructions and resources — no proprietary binary format
• Three trust tiers: system (built-in), curated (vetted), and experimental (use at your own risk)
• Install via $skill-installer inside Codex, either by name or raw GitHub URL
• Per-skill licensing inside each directory’s LICENSE.txt
• Open standard published separately at agentskills.io

Caveats

• The README is thin on what a skill actually looks like structurally; you’ll need to dig into the folders
• Experimental skills require more verbose install syntax than curated ones
• Restarting Codex after install is required, which feels slightly 1990s

Verdict

Worth browsing if you’re using Codex and tired of typing the same prompts. Skip it if you’re not in OpenAI’s agent ecosystem — the open standard is aspirational, but the tooling isn’t there yet for other platforms.