The Agent Skills Directory

[DYNAMIC_CONTEXT_INJECTION]: The SKILL.md file utilizes the dynamic context injection syntax (!) to execute a cat command when the skill is loaded. This is used for the legitimate purpose of inlining documentation from a related skill (craft) to provide context to the agent.
[COMMAND_EXECUTION]: The skill includes several internal scripts (check.sh, wc-probe.ts, parity-check.sh, validate-skill-write.ts) that execute shell commands and filesystem operations. These scripts are used to perform lints, run verification tests, and provide real-time feedback during file edits. The implementation includes safeguards such as path containment checks to prevent climbing out of the targeted audit directory.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to audit external directories and files provided by the user. This creates a potential surface where malicious instructions within an audited file could influence the behavior of the agent performing the audit.
Ingestion points: Files located within the directory targeted for auditing via the --target parameter in the provided scripts.
Boundary markers: The skill makes extensive use of Markdown code fences and specialized comment markers () to structure data and instructions.
Capability inventory: The skill has permissions for shell execution (Bash), filesystem modification (Write, Edit), and the ability to dispatch sub-agents.
Sanitization: The audit process relies on deterministic scanning logic (regular expressions and basic parsing) implemented in TypeScript rather than relying solely on LLM interpretation of the audited content.

workflow-creator