Skills

Use a skill as a focused teammate.

A skill is a reusable workflow: a name, a trigger description, and instructions, with optional scripts, references, assets, and UI metadata. It should change a decision or make a repeated task more reliable—not repeat generic advice.

Routing rule: choose the narrowest skill that genuinely matches the requested work. If a skill does not apply, do not load it just because it sounds adjacent.

Where skills live

Scope	Typical location	Use
Repository	`.agents/skills/<skill-name>/SKILL.md`	Shared project or subproject workflows that belong in version control.
User	`~/.agents/skills/`	Personal workflows useful across repositories.
System	Bundled with Codex	Broad workflows such as skill creation, installation, official docs, or reviews.
Plugin	Installed plugin package	Skills distributed with connectors or reusable product capabilities.
Admin	Machine-managed skills	Organization-wide defaults or controlled tool workflows.

In a repository, Codex scans relevant .agents/skills directories from the working directory upward to the repository root. If a skill changes but does not appear, restarting the Codex session can refresh discovery. See the official skill source for current details.

Catalog reviewed for this guide

This is a practical routing inventory of the skills available in this environment and the meta-skills reviewed from prior project work. Versioned plugin copies can change, so treat the source file and description as authoritative at task time.

Family	Skills	Use when…
Skill governance	`skill-creator`, `skill-installer`, `review-agent`, prior `agent-harness-review`	You are creating, installing, auditing, or reviewing skills and agent setup. Start with skill review.
Codex grounding	`openai-docs`, `deep-research`, `interview-me`	You need official OpenAI facts, an explicitly deep research artifact, or a one-question-at-a-time clarification interview.
Code and runtime	`review-agent`, `control-in-app-browser`, `control-chrome`, `computer-use`	You need a read-only code review, real local UI acceptance, existing browser state, or Windows app interaction.
Artifacts	`documents`, `pdf`, `Presentations`, `Spreadsheets`, `template-creator`	You are creating or verifying the named artifact type. Follow the artifact skill’s rendering or recalculation contract.
Visuals and explanation	`imagegen`, `visualize`	You need a bitmap asset or an interactive/static visualization. Do not use image generation to solve a code-native SVG/CSS problem.
Product surfaces	`sites-building`, `sites-hosting`, `plugin-management`	You explicitly want the Sites product, website hosting, or plugin discovery/management. Ordinary local HTML does not trigger Sites.

Current environment snapshot

As reviewed on September 13, 2026, this Codex session exposes these representative built-in or installed skills. This is a routing snapshot, not a promise that every machine has the same catalog.

Source	Observed skills
System / personal	`imagegen`, `openai-docs`, `plugin-creator`, `review-agent`, `skill-creator`, `skill-installer`, `interview-me`
Bundled tools	`control-in-app-browser`, `control-chrome`, `computer-use`, `visualize`
Curated workflows	`deep-research`, `plugin-management`, `sites-building`, `sites-hosting`, `sites-preview-troubleshooting`
Artifact workflows	`documents`, `pdf`, `Presentations`, `Spreadsheets`, `excel-live-control`, `template-creator`
Prior repository review	`agent-harness-review`, `repository-audit`, `verification-gate`, plus focused API, authorization, regression, browser, media, migration, performance, runtime, and UI review skills

Available is not installed everywhere. A recommended plugin list, a skill visible in a catalog, and a skill actually loaded for the current task are three different states. Confirm the current environment before relying on a name or path, and ask before installing anything.

Public skills worth investigating

This is a curated starting list, not an endorsement or an exhaustive marketplace. The links point to public source folders so you can inspect the trigger description, scripts, references, license, and maintenance before installing anything.

Skill / source	Use it for	Review note
`frontend-app-builder` OpenAI	Building a frontend app with visual assets and a coherent implementation loop.	Strong fit for new web work; read its surrounding plugin manifest and required tools.
`frontend-testing-debugging` OpenAI	Testing and debugging browser-facing frontend behavior.	Pair with a real browser check; a skill’s assertions are not a substitute for inspecting the result.
`react-best-practices` OpenAI	React and Next.js implementation decisions.	Use only when the project actually uses that stack; do not let framework advice override local conventions.
`remotion` OpenAI	Programmatic video and motion workflows in React.	Useful for media-as-code; verify render output, fonts, codecs, and licensing of assets.
`webapp-testing` Anthropic	Driving a local web app with Playwright, collecting screenshots, and checking browser logs.	Good acceptance-test pattern; keep it scoped to a local app and inspect any generated commands.
`frontend-design` Anthropic	Designing and implementing distinctive frontend interfaces.	Use for visual direction, then preserve the project’s accessibility and design-system constraints.
`algorithmic-art` Anthropic	Seeded generative art and interactive creative coding.	Useful for original browser-native media; inspect template requirements and third-party runtime links.
`vercel-react-best-practices` Vercel	React performance and framework-specific review.	Provider-specific advice needs a framework/version check and a measurable performance baseline.
`hf-cli` Hugging Face	Searching, downloading, uploading, and managing models, datasets, Spaces, and jobs.	Potentially powerful network and file operations; confirm destination, credentials, and license first.
`transformers-js` Hugging Face	Running machine-learning models in JavaScript or TypeScript.	Good for local/browser experiments; check model support, bundle size, and hardware constraints.
`systematic-debugging` Community	Evidence-led diagnosis before changing code.	Useful pattern, but community maintenance and permissions are your responsibility.
`verification-before-completion` Community	Checking claims against fresh command output before reporting completion.	Aligns with this handbook’s quality rules; still verify that its checks fit the project.

Directories and trust levels

OpenAI Plugins

Current official Codex plugin examples, including skill-backed web, media, mobile, and integration workflows. Prefer this current collection over the older deprecated OpenAI skills catalog.

Anthropic Skills

Official examples spanning development, browser testing, creative coding, documents, and enterprise workflows. Treat examples as inspectable patterns, not automatic guarantees.

Hugging Face Skills

Official Hub-focused skills for model, dataset, cloud, training, and JavaScript workflows; particularly relevant to media-model exploration.

skills.sh directory

A broad, dynamic discovery index. Use it to find candidates, then follow the source repository, license, audit signals, and recent maintenance before adoption.

Explicit

Call a skill by name

Use $skill-name in Codex or select the skill where the surface supports it. This is best when you know the workflow you want.

Implicit

Let the description route it

Codex may select a skill when the task matches its description. This is why the description needs a clear capability, trigger words, and meaningful exclusions.

Small caution about skill catalogs

Names and descriptions are selection metadata, not proof of quality. Read the full SKILL.md, inspect linked resources, and run the skill’s own validation or a realistic forward test before trusting it with important work.