Governed Vibecoding vs Unmanaged AI CodingRead Now →
Skip to main content

Codex Agent Skills

Learn how Codex Agent Skills use .agents/skills, progressive disclosure, plugins, optional metadata, and how they differ from AGENTS.md.

12 min read
Axiom Studio Team· Engineering

How Codex Agent Skills Work

Codex Agent Skills are official reusable workflow packages for Codex. They complement AGENTS.md: skills codify task-specific procedures, while AGENTS.md layers standing project instructions and repository context.

Source checked against OpenAI Codex skills docs

Official documentation source

OpenAI documents Agent Skills as reusable workflow packages for Codex CLI, IDE extension, and app.

Codex skills use progressive disclosure: initial context includes name, description, and path; full SKILL.md loads when selected.

Codex AGENTS.md is a separate custom-instruction mechanism that Codex reads before doing work.

Primary file

SKILL.md

Repo path

.agents/skills

Distribution

Direct folders for local authoring; plugins for reusable distribution

Related file

AGENTS.md for project instructions

Good fit

Repeatable Codex workflows such as reviews, migrations, docs updates, frontend checks, and operational runbooks.

Watch closely

Do not collapse AGENTS.md, plugins, MCP dependencies, and skills into one concept; each has a different job.

Governance move

Review .agents/skills changes, plugin bundles, optional scripts, and AGENTS.md overrides in the same approval trail.

OpenAI's Codex documentation describes Agent Skills as reusable workflows that package instructions, resources, and optional scripts. Skills are available in the Codex CLI, IDE extension, and Codex app.

Codex uses progressive disclosure. It starts with compact skill metadata, then reads the full SKILL.md when a skill is selected for the task.

Where Codex Skills Live

OpenAI Codex guide

Codex reads skills from repository, user, admin, and system locations. For repositories, it scans .agents/skills in directories from the current working directory up to the repository root.

Direct skill folders are best for local authoring and repo-scoped workflows. Plugins are the installable distribution unit when teams want to share reusable skills or bundle skills with app integrations.

1

$CWD/.agents/skills for skills relevant to the current working folder.

2

Parent .agents/skills folders for shared monorepo areas.

3

$REPO_ROOT/.agents/skills for repo-wide skills.

4

$HOME/.agents/skills for user-level skills.

5

/etc/codex/skills and bundled system skills for admin or OpenAI-provided capabilities.

Codex Skills vs AGENTS.md

OpenAI Codex guide

AGENTS.md is not the same thing as a Codex skill. Codex reads AGENTS.md before doing work so it has project instructions, working agreements, and context for the repository.

A skill is a reusable task package. Use AGENTS.md for standing rules that should always apply in a directory. Use a skill for a workflow that should load only when a matching task appears.

Do not over-port platform behavior

Each platform can use SKILL.md-style files, but discovery paths, invocation rules, frontmatter, permissions, and distribution mechanics differ. Verify the target platform before copying a skill unchanged.

Plugins and Optional Metadata

OpenAI Codex guide

Codex documentation positions plugins as the way to distribute reusable skills and apps beyond a single repo. Plugins can include skills, app mappings, MCP server configuration, and presentation assets.

Codex also documents optional agents/openai.yaml metadata for UI metadata, invocation policy, and tool dependencies such as MCP servers.

Permissions and Security

OpenAI Codex guide

Codex skills can include optional scripts and references, and plugins can declare tool dependencies. Treat those additions like production automation, not passive documentation.

Teams should review skill folders, plugin manifests, MCP dependencies, and AGENTS.md files together because all of them can steer Codex's behavior.

1

Keep skill descriptions precise so implicit matching does not trigger too broadly.

2

Prefer instructions over scripts unless deterministic behavior or external tooling is required.

3

Disable implicit invocation when a skill should be used only by explicit request.

4

Audit plugins before installing them across a team.

Enterprise Governance Checklist

OpenAI Codex guide

The Codex governance model should answer which AGENTS.md files were loaded, which skill was selected, whether a plugin contributed it, and which tools or MCP dependencies were available.

VibeFlow-style work tracking pairs well with Codex skills because each reusable workflow can be tied to a work item, execution log, commit, security review, and QA gate.

Govern Codex Skills with tracked agent work

VibeFlow connects reusable agent workflows to work items, execution logs, commit records, security review, QA, and durable project context. That audit trail makes skills reviewable instead of invisible prompt behavior.

See VibeFlow

Ready to get started?

See how Axiom Studio can transform your AI infrastructure with enterprise-grade governance, security, and cost optimization.

Contact Us