Skill Creator

Create skills that give Codex useful, non-obvious guidance without constraining unrelated work.

Core Principles

Assume Codex is already capable. Include only information that changes its decisions or improves its work. Remove generic advice, repeated instructions, speculative edge cases, and examples that do not materially clarify the task.

Preserve user intent and scope. A skill should support the requested task, not replace the user's chosen product, expand the assignment, modify unrelated configuration, or imply permission for additional external actions. Do not turn a particular example, past failure, or personal preference into a universal requirement.

Approval to complete a task does not expand its scope or execution permissions. For retrying or externally mutating workflows, define a stopping condition proportional to the risk.

Match specificity to the risk. Give the model room to choose an appropriate approach when multiple approaches are reasonable. Use detailed steps, deterministic scripts, or absolute language only when correctness, safety, permissions, or a genuinely fragile workflow requires them.

For open-ended work, describe the outcome and relevant decision criteria. For workflows with a preferred shape, offer useful examples or configurable scripts. Reserve fixed sequences and narrow parameters for operations where deviation would cause a concrete problem. Preserve non-obvious operational invariants, distinguish actual requirements from optional recommendations or local conventions, and avoid restating policies already enforced elsewhere.

Keep discovery cheap and precise. Skill names and descriptions are available before a skill is loaded. Describe the actual capability and when it applies, adding exclusions only when they prevent likely misrouting. Avoid exhaustive capability lists and catchalls that attract unrelated requests.