| name | skill-creator |
| description | Guide for creating or updating reusable portable Agent Skills. Use when the user wants to design, create, refine, validate, or forward-test a Skill package without imposing a host-specific authoring gate. |
Skill Creator
Guide the creation or refinement of reusable, portable Agent Skills. A Skill is a focused package of instructions and optional resources that helps an Agent perform a recurring class of work consistently.
This guidance does not own filesystem access, permissions, activation, or trust. Any host-supported authoring path may create the same portable package. Choose the available path that best fits the task and the user's instructions.
Decide Before Writing
- Confirm that reusable guidance is more appropriate than a one-off answer or a product feature.
- Identify the recurring trigger, expected outcome, important constraints, and evidence of success.
- Choose the smallest useful package: instructions only, or instructions plus reusable scripts, references, assets, or host metadata.
- Keep the portable method in the Skill and keep host runtime protocols, internal schemas, and product-specific wiring outside it.
Use a short lowercase hyphenated name. Write a description that states both what the Skill does and when it should be used.
Portable Package
A portable Skill package has this shape:
skill-name/
├── SKILL.md
├── scripts/ # optional
├── references/ # optional
├── assets/ # optional
└── agents/
└── <host>.yaml # optional host metadata
SKILL.md is required. Its frontmatter defines name and description; its body contains the guidance loaded when the Skill is used.scripts/ contains deterministic or frequently repeated operations.references/ contains detailed material that should be loaded only when needed.assets/ contains templates or files used in produced output rather than prompt context.agents/<host>.yaml is an optional host overlay. Preserve overlays for other hosts when updating a shared Skill.- A root
manifest.json is not part of the portable Skill contract. Do not require one for creation or reuse.
Authoring Principles
- Be concise. Assume the Agent already knows general concepts and include only task-specific judgment, procedure, or constraints.
- Match specificity to risk: use flexible guidance for contextual work and scripts or exact steps for fragile deterministic work.
- Put the main workflow in
SKILL.md; move large examples, schemas, and background material into references/.
- Make references discoverable from
SKILL.md and avoid deep chains of references.
- Reuse existing files when updating a Skill. Do not replace user-authored resources or other-host overlays without a task-specific reason.
- Keep secrets, machine-specific absolute paths, runtime tool schemas, and host-internal protocols out of portable content.
Draft, review, and apply may be useful authoring techniques, but they are not mandatory Skill-creation gates. User approval is required only when the active host policy or the user's instructions require it; the Skill itself must not invent an approval barrier.
Validation
Before reporting completion:
- Check the package shape, frontmatter, names, links, and referenced files.
- Run bundled scripts or focused tests when present.
- Inspect the Skill for duplicated generic knowledge, hidden host coupling, obsolete manifest requirements, and unsupported claims.
- Forward-test with realistic prompts that should trigger the Skill and nearby prompts that should not.
- Report what was created or changed, what was validated, and any remaining uncertainty. Do not claim files were written unless the selected authoring path confirmed the write.
