Agent Skills · Browse the collection · Contribute · MIT licensed

Focused, portable instructions for Codex, Claude Code, and other coding agents that support the Agent Skills format. Each skill does one job, keeps its behavior in a standard SKILL.md, and includes host metadata only where a host needs it.

Start with the catalog

Inspect every available portable skill before installing anything:

npx skills add Emlembow/skills --list

Project scope is the default and is usually the safest choice for a team repository. This collection is installed from GitHub with npx skills add Emlembow/skills; it is not an npm package.

Choose a skill

Skill	Best for
research-loop	Metric-driven implementation experiments protected by holdout and leakage checks
adversarial-review	Two independent attempts to disprove a versioned, digest-checked result

adversarial-review is also indexed on skills.sh. GitHub-hosted skills appear there after an install through the skills CLI with anonymous telemetry enabled.

Install exactly what you need

Every portable skill can be selected independently:

npx skills add Emlembow/skills --skill research-loop
npx skills add Emlembow/skills --skill adversarial-review

To install a reviewed skill for a specific agent at user scope, be explicit:

npx skills add Emlembow/skills --skill research-loop --agent codex --global --yes
npx skills add Emlembow/skills --skill research-loop --agent claude-code --global --yes

The CLI recommends symlink installation. Use --copy only when the target environment cannot use symlinks. Update project-scoped skills with npx skills update -p, or user-scoped skills with npx skills update -g.

Review before granting access. Skills can contain executable scripts. Read a skill and its bundled resources before giving it access to sensitive repositories, credentials, or external systems. Set DISABLE_TELEMETRY=1 or DO_NOT_TRACK=1 when telemetry must be disabled.

Portable at the core

The repository keeps reusable behavior separate from host presentation and distribution:

skills/<skill-name>/
├── SKILL.md              # portable behavior and trigger rules
├── agents/openai.yaml    # Codex display metadata
├── references/           # optional focused detail
├── scripts/              # deterministic helpers, only when needed
└── .claude-plugin/       # Claude plugin metadata, when distributed there

• The top-level skills/ folders are the source discovered by npx skills.
• agents/openai.yaml adds Codex presentation without changing portable behavior.
• Claude plugin manifests and marketplaces point back to the same skill folders.
• Repository validation checks structure, metadata, discovery, and marketplace paths together.

Create or contribute a skill

Use the creator built into your agent when available ($skill-creator in Codex), or initialize a portable skill with:

npx skills init my-skill

Keep SKILL.md focused on one job. Put trigger and non-trigger conditions in its frontmatter description, write imperative steps with explicit inputs and outputs, and move optional detail into one-level-deep references/. Add agents/openai.yaml for Codex presentation metadata.

See CONTRIBUTING.md for the complete repository requirements.

Validate the collection

The automated checks pin tool versions for reproducibility even though end-user examples use the official unversioned npx skills form.

npm run validate

This checks repository structure, skill metadata, top-level portable discovery, and the Claude marketplace. Pull requests and pushes to main run the same validation in GitHub Actions.

References

• OpenAI: Build skills
• Agent Skills specification
• skills CLI reference
• Claude Code skills

License

MIT. See LICENSE.