The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to modify local files within the src/skills/ directory to apply fixes and improvements found during the audit process.
[PROMPT_INJECTION]: The skill ingests untrusted markdown data from target skills for analysis. This creates a potential surface for indirect prompt injection, where malicious instructions within a skill being audited could attempt to influence the auditor's judgment or findings.
Ingestion points: The auditor reads the target skill's SKILL.md and all files under its references/ directory.
Boundary markers: The skill uses a structured "unit ledger" to decompose and categorize target content into typed units (directives, examples, data-points).
Capability inventory: The skill performs file system read and write operations within the src/skills/ directory and suggests the use of external tools like bin/build and md-audit.
Sanitization: The skill relies on its multi-pass structured analysis and mandatory human confirmation for contradiction resolution rather than explicit sanitization of the input text.

skill-audit