| name | skill-creator |
| description | Use when creating, packaging, reviewing, or improving an Agent Skill for Codex, Claude, Trae, or another SKILL.md-compatible runtime. |
Skill Creator
Turn repeatable expertise into a small, discoverable Skill rather than a one-off prompt.
Define the contract
Before writing files, establish:
- Which concrete user situations should trigger the Skill.
- The expected result and what counts as success.
- Required tools, permissions, dependencies, and supported runtimes.
- Dangerous actions that require explicit confirmation.
- Two realistic positive examples and one nearby case that should not trigger.
Package
Create a kebab-case directory containing SKILL.md. Put only name and a trigger-focused description in required frontmatter. Keep the main instructions concise and imperative. Move large references into references/, deterministic operations into scripts/, and reusable templates into assets/.
Do not assume a tool exists because another runtime provides it. Describe capabilities and prerequisites explicitly. Never embed secrets, machine-specific absolute paths, or instructions that silently transmit or delete user data.
Validate
- Confirm the name matches the directory and uses only letters, numbers, and hyphens.
- Verify frontmatter parses and referenced files exist.
- Run realistic prompts with and without the Skill when the environment supports evaluation.
- Check that the Skill improves the intended result, does not trigger on unrelated requests, and degrades clearly when dependencies are unavailable.
- Keep changes small; revise instructions based on observed failures rather than adding broad rules speculatively.
Install target
When the user requests direct installation, choose the runtime explicitly: ~/.codex/skills/<name>, ~/.claude/skills/<name>, or ~/.trae/skills/<name>. Refuse to overwrite an unrelated existing directory without confirmation. For this application, user-created Skills remain user resources and must not be written into the bundled official catalog.
