OpenAI Skills

Executive Summary

OpenAI Skills is the official skills catalog for Codex, OpenAI's coding agent. It defines a portable format — Agent Skills — where each skill is a folder containing instructions, scripts, and resources that an AI agent can discover and execute. As of June 2026 the repo has ~22,000 GitHub stars and ~1,500 forks (up from ~9,000 stars when this profile was first written in February), and serves as both a distribution mechanism and a reference implementation of the open Agent Skills standard. ^{[1] [2]}

How It Works

The catalog organizes skills into three tiers ^[1]:

• System (.system/) — Auto-installed in the latest Codex version
• Curated (.curated/) — Vetted skills, installable by name (e.g. $skill-installer gh-address-comments)
• Experimental (.experimental/) — Community and exploratory skills, installed by folder path or GitHub URL

On the Codex side, skills now resolve across four scopes: repository (.agents/skills), user (~/.agents/skills), admin (/etc/codex/skills), and built-in system skills bundled with Codex. ^[3] Codex activates skills explicitly (a $skill mention) or implicitly when the task matches, using progressive disclosure — only skill metadata is loaded up front (capped around 8,000 characters) until a skill is selected, when the full SKILL.md is read. Installed skills are detected automatically; restarting Codex resolves detection issues. Skills can be disabled without deletion via ~/.codex/config.toml, and packaged as plugins with optional app mappings and MCP configurations. ^{[3] [4]}

What's Changed Since February 2026

• Star growth ~9k → ~22k — the catalog more than doubled its GitHub traction in roughly four months. ^[1]
• Cross-agent adoption of the standard — Microsoft adopted Agent Skills in VS Code and GitHub Copilot, as did Cursor, Goose, Amp, and OpenCode; the SKILL.md format that originated with Anthropic's Claude Skills now spans the major coding agents. ^{[2] [5]}
• Richer runtime model — four-scope resolution, progressive disclosure, implicit activation, config-based disabling, and plugin packaging with agents/openai.yaml metadata. ^[3]
• Symphony (April 27, 2026) — OpenAI open-sourced a spec for Codex orchestration that turns an issue tracker like Linear into a control plane: every open task gets an agent, agents run continuously in separate workspaces and monitor CI, and humans review results. It ships as a SPEC.md plus an Elixir/BEAM reference implementation under Apache 2.0. OpenAI says internal teams saw landed pull requests rise 500% in the first three weeks. ^{[6] [7]}

Why It Matters

The "write once, use everywhere" framing is no longer aspirational — the Agent Skills standard has real multi-vendor adoption, making skills the closest thing the agent ecosystem has to a package format. ^[2] Symphony extends the same philosophy upward: if skills are the unit of agent capability, Symphony is the spec for coordinating many capable agents at once. ^[6]

Competitive Context

This is comparable to:

• Claude Code's Agent Skills — Anthropic originated the SKILL.md format; the two ecosystems now share a standard ^[5]
• GitHub Copilot / VS Code agent skills support ^[2]
• Cursor's rules and skills support

The differentiator has shifted: with the format itself commoditized across vendors, OpenAI's edge is the curated catalog, the tiered trust model, and the Symphony orchestration layer above it.

Cautions

• Free-form Markdown is hard to verify — skills are unrestricted Markdown bodies behind YAML frontmatter, which makes them easy to author but difficult to evaluate or audit systematically. ^[5]
• No secrets mechanism — the spec offers no standard way to inject credentials into skill scripts, forcing per-execution prompts or hardcoding. ^[5]
• Symphony is a spec, not a product — the repository is essentially a SPEC.md with a reference implementation; the operational burden of running agent fleets remains yours. ^[7]
• Per-skill licensing — there is no single repo license; each skill carries its own LICENSE.txt. ^[1]

What Developers Say

The Hacker News launch thread for skills in Codex drew 304 points, with sentiment largely positive on the standard and skeptical on verifiability. ^[5]

"It's so nice that skills are becoming a standard, they are imo a much bigger deal long-term than e.g. MCP." — cube2222, Hacker News ^[5]

"At my startup, we have a mix of Codex/CC users so having a common set of skills we can all use for building is exciting." — rdli, Hacker News ^[5]

"Are we sure that unrestricted free-form Markdown content is the best configuration format for this kind of thing? ... doesn't the free-form nature of the 'body' part of these configuration files lead to an inevitably unverifiable process?" — mikaelaast, Hacker News ^[5]

On Symphony, the reaction split between strategy-watching and genuine enthusiasm:

"OpenAI is doing the Joel Spolsky move, commoditize the layer above yours. ... The spec is free, the runtime is where the money is." — 7777777phil, Hacker News ^[8]

"This is cool. I've been waiting for something like this! ... This might just make me take another look at Codex, though I'm sure it'll be adaptable to other harnesses." — solarkraft, Hacker News ^[8]

Bottom Line

OpenAI Skills graduated from a promising catalog to the reference distribution channel for a genuinely cross-vendor standard: ~22k stars, four-scope runtime resolution, and SKILL.md support spanning Codex, Claude Code, Copilot, Cursor, Goose, Amp, and OpenCode. Symphony adds an orchestration layer that turns the skills bet into a fleet-management story.

Recommended for: Teams standardizing agent capabilities across mixed Codex/Claude Code shops, and anyone packaging repeatable workflows for coding agents.

Not recommended for: Workflows needing verifiable, auditable agent configuration or built-in secrets handling — the free-form Markdown format has no answer for either yet.

Outlook: Watch whether Symphony implementations materialize beyond OpenAI's Elixir reference and whether the standard adds evaluation and secrets primitives; both would cement skills as the package manager of the agent era.