---

name: skill-creator description: Create, update, or refine custom Grok skills. Use when the user wants to build a new skill, improve an existing one, design skill structure or format, add scripts/references/assets, make a skill more autonomous, package domain knowledge, polish a SKILL.md, or deploy a skill to Grok bot. Triggers on create skill, new skill, make a skill, skill creator, update this skill, skill format, skill template, design skill for, build custom skill, how do I make a skill, improve skill triggers, deploy skill to grok bot, or any request to author or polish a SKILL.md. This skill supports autonomous activation and full ecosystem testing when coordinated via skill-orchestrator. metadata: short-description: Author, polish, and package Grok skills argument-hint: "[skill-name or polish target]" surfaces: grok-bot, grok-chat, grok-ios, grok-web, grok-build, grok-cli slash: /skill-creator

Skill Creator

Create and manage skills — modular instruction packages that specialize the agent. Encode non-obvious, procedural, or organization-specific knowledge the base model does not have.

Core rule. Only create a skill if the knowledge justifies its token cost. Do not duplicate base model capabilities. Encode workflows, templates, locked standards, domain procedures, and integration patterns.

Skill structure

skill-name/
├── SKILL.md         # Required — frontmatter + imperative instructions
├── scripts/         # Optional — deterministic work outside context
├── references/      # Optional — loaded on demand
└── assets/          # Optional — copied, not read

Frontmatter (strict — PKE CI)

Allowed top-level keys only: name, description, license, compatibility, metadata, allowed-tools.

Grok also understands user-invocable, when-to-use, argument-hint, paths. Put those under metadata: so validate-skill.sh stays green.

• name — kebab-case, matches directory, 2–64 chars, start/end alphanumeric, no consecutive hyphens, unquoted.
• description — what + when + synonyms. Plain YAML scalar. Max 1024 chars. No colon-space, no quotes, no angle brackets, no TODO.

Body

Imperative. Challenge every sentence: does this justify tokens?

• Keep lean. At ~350–400 lines, move detail to references/.
• Link references relatively. Load only when needed.
• Cross-link skill-orchestrator and skill-test-suite.
• Portable paths. Never hardcode /root/.grok/server-skills. Resolve in order: $PKE_ROOT, ~/.grok/skills, /workspace/.grok/skills, repo root.

Perfect-a-skill loop

1. Recall — read the existing SKILL.md + references + scripts.
2. Audit — run validate-skill.sh. Score triggers, length, hardcoded paths, missing orchestrator link, duplicate see-also, missing metadata.
3. Improve — broaden triggers, lean the body, split references, add Grok-bot slash + surfaces in metadata, fix paths.
4. Validate — bash "$CREATOR/scripts/validate-skill.sh" "<dir>" until OK.
5. Integrate — trigger skill-orchestrator (inventory, dep graph, agent meeting if the skill sits on a sequence).
6. Deploy — symlink ~/.grok/skills/<name>, copy into Build .grok/skills/, paste to grok.com custom skills. Slash /<name>.
7. Automate — if it should run daily, add a Grok automation whose prompt names the skill.

Init (portable):

CREATOR="$(dirname "$0")/.."   # or $PKE_ROOT/skill-creator
bash "$CREATOR/scripts/init-skill.sh" <skill-name> "$PKE_ROOT"
bash "$CREATOR/scripts/validate-skill.sh" "$PKE_ROOT/<skill-name>"

Quality rubric

| Gate | Pass | | --- | --- | | CI frontmatter | validate-skill.sh OK | | Triggers | what + when + ≥3 natural phrasings | | Lean | body < 350 lines or split with locked blocks kept | | Portable | no /root/.grok hardcodes | | Ecosystem | orchestrator + test-suite mentioned | | Bot | metadata.slash + surfaces | | Honest | no claim of weight/quota changes |

Domain patterns: references/domain-patterns.md. Design patterns: references/design-patterns.md. Troubleshooting: references/troubleshooting.md.

Pretty Kitty examples: paralegal recovery, brand voice, locked visual identity, media chains, testing and orchestration.

See also: skill-orchestrator, skill-test-suite, autonomous-ecosystem, PKEMEDIA/pke-ai-agent-skills.