The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process and audit instruction files from project or global directories, creating a surface for indirect prompt injection. The skill mitigates this by instructing the agent to treat commands in audited files only as evidence and by using a specialized script for static analysis.
Ingestion points: The skill reads instruction files like skill.md, agents.md, and claude.md for review.
Boundary markers: SKILL.md explicitly states that commands found in the objects being audited are not to be executed and are only used as evidence.
Capability inventory: Uses a bundled Python script (inventory.py) for read-only metadata collection; no network operations are present in the skill's code.
Sanitization: The analysis script utilizes yaml.safe_load() for frontmatter parsing and handles missing parser dependencies safely.
[COMMAND_EXECUTION]: The skill utilizes a localized Python script to build a file inventory. The script implements security best practices by using safe loading methods, detecting symlink cycles, and preventing the overwriting of existing files through strict file creation checks for its output.

audit-agents-md