skill-sanitize

What problem does it solve?

When importing third-party skills, copying skills from reference repositories, or bulk-migrating skills between kits, there is a risk of hidden model-specific delimiter tokens that enable prompt injection attacks, or overly long frontmatter/body fields that break runtime ingestion. This Skill eliminates that risk by running a pre-load safety scan before any untrusted skill content is loaded into your environment.

Core Features & Use Cases

• Delimiter Token Scanning: Detects high-risk model-specific tokens (ChatML, Llama, Llama 3 delimiters) in skill frontmatter and body content.
• Length Cap Validation: Checks skill description and instruction fields against upstream and kit-specific length limits to prevent runtime failures.
• Opt-In Natural Language Scanning: Supports extended pattern matching for natural language prompt injection attempts for high-stakes untrusted source ingestion.
• Use Case: A team importing 20+ skills from a public marketplace can run this scan in bulk to flag any skills with hidden injection risks or non-compliant field lengths before adding them to their kit.

Quick Start

Use the skill-sanitize skill to scan the .claude/skills/ directory for unsafe delimiter tokens and length violations.