OpenAI Launches Skills in Codex to Customize Coding Agents

OpenAI Codex Skills, introduced in December 2025, let developers customize coding agents with reusable SKILL.md “skills” that standardize workflows across Codex CLI and the Codex IDE extension.​

What OpenAI launched

OpenAI’s new Agent Skills feature extends Codex with task-specific capabilities packaged as a “skill,” which can include instructions plus optional scripts and supporting resources.​

OpenAI says skills follow an open Agent Skills standard and use “progressive disclosure,” meaning Codex loads only each skill’s name and description at startup, then reads the full instructions only when the skill is used.​

Skills are available in Codex CLI and IDE extensions, with explicit skill calling (typing $ or using /skills) supported in those surfaces, while repository-checked skills can still influence Codex behavior even where explicit invocation is not supported.​

What a “skill” contains

A skill is centered on a SKILL.md file with YAML front matter (including required name and description) and an optional instruction body, and it may also include folders like scripts/, references/, and assets/ for workflows that need deterministic steps, templates, or extra documentation.​

OpenAI’s documentation emphasizes that Codex injects only the skill name, description, and file path into runtime context by default, keeping the full body on disk unless the skill is invoked.​

How it fits Codex’s evolution

OpenAI launched Codex in May 2025 as a cloud-based software engineering agent that can work on multiple tasks in parallel in isolated sandbox environments preloaded with a repository.​

OpenAI describes Codex as capable of writing features, answering questions about a codebase, fixing bugs, and proposing pull requests, with evidence such as terminal logs and test outputs to help users verify what the agent did.​

In September 2025, OpenAI announced upgrades including GPT‑5‑Codex (optimized for agentic coding), broader product unification across terminal/IDE/web/GitHub/iOS, and expanded code review workflows designed to catch critical issues earlier in PR review.​

Why OpenAI Codex Skills matter

OpenAI positions skills as a way to capture institutional knowledge so Codex behaves consistently across users, repositories, and sessions—especially for teams that want repeatable engineering standards.​

Because Codex can select skills implicitly when a task matches a skill’s description, teams can encode “when to apply this workflow” directly into the system, rather than relying on every developer to remember long prompt templates.​

This approach also aims to keep context smaller and more predictable, since Codex does not automatically load entire skill documents into the prompt unless needed.​

Examples of workflows teams can turn into skills include:

• Code review checklists and conventions that should apply on every PR.​
• Security or compliance checks that should run before code is submitted.​
• Repeatable analysis tasks (for example, structured incident triage or dependency audits).​
• Team-specific templates (commit messages, release notes, design-to-implementation checklists).​

How skills work in practice

Codex can activate skills either explicitly (the user names or selects the skill) or implicitly (Codex decides the skill applies based on the description).​

OpenAI documents built-in system skills including $skill-creator (to scaffold new skills) and $skill-installer (to install skills from curated sources), plus an experimental $create-plan skill that must be installed.​

OpenAI’s Codex changelog also describes installing skills per-user (for example under ~/.codex/skills) or per-repo (checked into .codex/skills in a repository) to standardize behavior across a project.​

Skill storage and precedence

OpenAI documents multiple skill scopes and locations, with repo-scoped skill folders taking precedence over user/admin/system layers when names collide.​

Skill creation options

OpenAI supports creating skills with $skill-creator (guided scaffolding) or manually by creating a skill directory and a SKILL.md file in a supported path.​

The docs recommend using “instruction-only” skills by default and adding scripts only when determinism or integrations are needed.​

OpenAI also stresses that clear triggers matter because the description is the primary signal used for automatic triggering.​

Enabling and configuring

OpenAI’s Codex configuration documentation lists skills as a feature flag and notes it can be enabled via configuration or CLI flags (for example, codex –enable feature_name, including multiple flags).​

Community discussion around the rollout has also referenced skills being gated as an experiment in some setups, aligning with the platform’s broader pattern of staged feature releases.​

Timeline: from Codex to Skills

OpenAI first introduced Codex as a cloud agent in May 2025 and continued shipping major workflow upgrades through September 2025, before adding skills support in December 2025.​

Final thoughts

OpenAI Codex Skills push coding agents toward a more “configurable teammate” model, where workflows can be packaged, shared, and reused instead of rewritten in every prompt.​

For engineering leaders, skills create a practical mechanism to standardize reviews, testing steps, and security checks inside the same agent loop Codex already uses for editing and verification.​

For developers, the biggest immediate impact is portability of best practices across repos and teams—without forcing skill bodies into every session’s context.​