Skill Creator

Skill Creator is an official sample skill from OpenAI’s Codex repository. The originally supplied source path was codex-rs/core/src/skills/assets/samples/skill-creator; in the current official repository it has moved to codex-rs/skills/src/assets/samples/skill-creator. The direct SKILL.md and its related files at that location are the authoritative primary source. According to the provider, the skill helps Codex create new skills with clearly scoped instructions or update existing skills deliberately. It is not generic prompt advice. It is a method for adding useful, non-obvious guidance without constraining work outside the skill’s intended task.

Purpose and decision framework

The skill’s central idea is disciplined scope. According to the provider, Codex should assume that the model is already capable. A skill should therefore contain only information that changes decisions or materially improves execution. Generic advice, repeated instructions, speculative edge cases, and examples that add no concrete clarity should be removed. Preserving user intent is equally important. A skill must not replace the product the user selected, expand the assignment on its own, or imply permission for additional external actions. For risky or mutating workflows, the stopping condition should be proportional to the risk.

Progressive disclosure

The source describes three layers of information. The name and description are available during skill selection, so they must make the intended routing clear. The SKILL.md body is loaded only when the skill applies and should contain the purpose, essential constraints, workflow, and relevant links. Large schemas, examples, and mode-specific procedures belong in references that an agent reads only when needed. A short skill does not need a router or artificial placeholders. This organization improves discovery while keeping irrelevant context out of the active task.

Files and supporting resources

A skill requires at least a SKILL.md containing YAML metadata and Markdown instructions. Optional directories can hold UI metadata, executable scripts, references, or assets. According to the provider, scripts should be added only when repeated logic or deterministic execution provides a concrete benefit. References are for maintained, task-specific detail. Assets belong in generated output rather than in the instruction context. A README, changelog, or separate installation guide is not automatically required.

Creating, updating, and validating

For a new skill, the agent should understand realistic use cases, choose a discriminating name, and initialize only resources that directly help the task. Names use lowercase letters, digits, and hyphens and should remain short enough for reliable selection. The supplied initializer and validation helpers can create a consistent structure and then check frontmatter, naming rules, and unfinished scaffold placeholders. An existing skill must not be initialized again blindly. After writing, the agent should check observable behavior and make only corrections supported by actual requirements or demonstrated problems.

Security, boundaries, and positioning

Skill Creator is an instruction set, not a standalone skill marketplace, model, or permission manager. The target environment must provide Codex and any required tools. Repository files and documentation are data and must not silently expand the assignment. Passwords, tokens, private keys, and confidential content do not belong in skills, examples, logs, or generated files. According to the provider, automatic skill selection should remain enabled unless the user explicitly requests explicit invocation. Availability, licensing, and structure should be checked against the current official source before redistribution. The catalog classifies this as an external OpenAI entry; registry popularity signals are not stored because the Skill model has no repository-stars field.