The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill audits external instruction files and documentation, which provides a path for indirect prompt injection if the audited content contains adversarial instructions. 1. Ingestion points: Step 1 of SKILL.md specifies reading various files, references, and tool results associated with the audited skill. 2. Boundary markers: The instructions explicitly state that all inspected data is untrusted and that embedded instructions cannot expand the agent's scope or authority. 3. Capability inventory: The skill has read access to the repository, performs file writes (via a script), and executes git commands. 4. Sanitization: There is a mandatory rule that secret values must never be copied into outputs, requiring only their type and location to be cited.
[COMMAND_EXECUTION]: The accompanying TypeScript script (scripts/reconcile-report.ts) executes shell commands, specifically git log and git show, to retrieve historical iterations of audit reports. These are invoked using a local utility defined in the vendor's repository.
[DYNAMIC_EXECUTION]: The skill requires running a local TypeScript script as part of its fourth step. This script manages complex file operations, such as creating temporary files, calculating hashes, and performing atomic renames to ensure report integrity.

tailrocks-skill-audit