About Codex Skills
An independent catalog of skills for OpenAI Codex, written to one standard: specific enough that the agent works like a careful senior practitioner, and open enough that you can read every line before you install it.
Why this exists
Skills are the cheapest way to give Codex durable expertise. A short description sits in context, and the full procedure loads only when a task needs it. Most skills shared so far are either narrow one-offs or a few paragraphs of advice Codex would follow anyway. codexskills.dev collects 872 skills that try to write down what an experienced practitioner actually does: the order of operations, the thresholds, the traps and the check at the end.
Editorial standards
Every skill is held to the same written specification before it is published:
- A selective description. The frontmatter says what the skill does, with concrete tools and file types, and names the requests that should load it. That is the only text Codex sees before deciding, so vague descriptions are rejected.
- Procedure over prose. A numbered workflow, decision tables for the real forks, and exact commands and configuration for current, stable tool versions.
- Written for the sandbox. Codex often runs with restricted network and file access. Skills prefer commands that work offline, say when a step needs network or wider access, and never tell Codex to get around an approval.
- Pitfalls and verification. Each skill names the mistakes that cause real incidents and ends with how to prove the work is correct.
- Safety. Destructive steps require explicit confirmation inside the workflow. Security skills cover defensive work and testing you are authorized to do, nothing else.
- Portability. Only the standard frontmatter fields (
name,description,license) are used, following the Agent Skills specification. - Nothing hidden. Every page renders the complete
SKILL.mdand every bundled script or reference. The zip holds exactly those files plus a generatedagents/openai.yamlfor the Codex app.
How skills are written and checked
Each skill starts from a scoped brief so that neighbouring skills hand off instead of overlapping. Drafts are written with AI assistance and then edited against the specification above. An automated validator checks the frontmatter, the description length and wording, references to bundled files and to sibling skills, and the metadata shown on the site. Bundled scripts are kept small and dependency-light so they can be read in a minute.
Tools and APIs change. When a command or version in a skill goes stale, report it and it will be corrected. Corrections take priority over new skills.
Where installed skills go
The installer and CLI write personal skills to ~/.agents/skills/ and project skills to
.agents/skills/, the folders Codex scans in the CLI, the IDE extension and the Codex app. Nothing else on
your machine is touched, and the scripts send no telemetry. Read install.sh.
Licensing
Every skill is released under the MIT license. Use, change and redistribute them, including inside commercial products and private plugin collections, as long as the license notice stays.
Corrections and contact
Found an error, a broken command or a skill that loads when it should not? Email [email protected] with the skill name and what you expected to happen.
Independence
Codex Skills is an independent, community-run directory. It is not affiliated with, endorsed by, or sponsored by OpenAI. "OpenAI" and "Codex" are trademarks of OpenAI, used here only to describe the tool these skills are written for. For how Codex itself behaves, the official Codex skills documentation is the reference.
The same editors run sister directories for other agents: claudeskills.page for Claude Code and opencodeskills.dev for OpenCode.