skill-creator

# Skill Creator Inspect the destination's conventions and existing skill before editing. When available, load `skills-assist` for current Amplifier authoring guidance. Derive the workflow from concrete user tasks; include instructions that change useful decisions rather than repeating generic assistant behavior. Use a unique lowercase hyphenated `name` and a concise `description` explaining when to load the skill. Add `user-invocable: true` when a slash command is useful. Match the name to its directory, use at most 64 characters with no consecutive hyphens, and keep the description within the standard's 1024-character limit and the destination's tighter routing budget. Use string `compatibility` (up to 500 characters) for actual environment requirements; keep `metadata` values strings. `version`, `context`, and `user-invocable` are host extensions, not standard Agent Skills fields. See https://agentskills.io/specification. Keep interactive work inline. Use `context: fork` only for a bounded independent task, and only with the current host's support. Fork `allowed-tools` uses module IDs; omit it when the full parent tool surface is appropriate. Do not pin a provider/model unless required by the user's requested capability. Put essential steps and completion evidence in `SKILL.md`. Store conditional detail in linked references, deterministic helpers in scripts, and reusable output inputs in assets. Resolve resources relative to `skill_directory` or `${SKILL_DIR}`; do not assume the caller's working directory. Save to the authorized bundle or `.amplifier/skills/<name>` destination. Preserve unrelated files and notices. Parse frontmatter, verify linked resources, load through the actual `load_skill` tool, and test a realistic scenario including missing optional capabilities. Report loader, execution, and visual checks separately. Update an existing skill in place only when that target is requested. Test descriptions with ordinary requests and near misses using the full catalog; explicitly naming a skill does not test selection. Record actual loads and independent output checks in fresh sessions. For a quality-improvement claim, compare against the previous version or a no-skill baseline with the same tools and model. See https://agentskills.io/skill-creation/evaluating-skills.