The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/count-context.py script invokes local git binaries via subprocess.run to perform repository tree listing and file content extraction for version comparisons. Arguments are passed as a list, which prevents shell injection.
[EXTERNAL_DOWNLOADS]: The skill instructions specify requirements for tiktoken and PyYAML, which are standard libraries for tokenization and metadata parsing. It also references a skills CLI tool intended for package management.
[INDIRECT_PROMPT_INJECTION]: The auditing tools read and process external SKILL.md files, which could contain malicious instructions designed to influence the auditor's behavior.
Ingestion points: scripts/audit-package.py and scripts/count-context.py read the content of SKILL.md and referenced files from target skill directories.
Boundary markers: The auditor utilizes yaml.safe_load to isolate frontmatter metadata from instructional content and uses standard markdown parsing for links.
Capability inventory: The scripts perform read-only static analysis and output results to the console. They do not implement any logic to execute commands found within the audited skill content.
Sanitization: The audit-package.py script includes explicit validation logic to ensure that local links and symlinks do not resolve to locations outside the audited package's root directory.

agent-skills-standard