The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external SKILL.md files and modules, which creates an inherent surface for indirect prompt injection. 1. Ingestion points: Tools such as skills-auditor and token-usage-tracker in the scripts directory read content from user-provided skill files and the ~/.claude/skills/ directory. 2. Boundary markers: The provided files do not specify explicit delimiters or instructions to ignore embedded commands when processing the external skill content. 3. Capability inventory: The framework utilizes subprocess.run to execute analysis scripts and performs direct file reads via Path.read_text. 4. Sanitization: There is no evidence of sanitization or escaping for the ingested content before it is processed or presented in reports.
[COMMAND_EXECUTION]: The skill framework executes local analysis tools and scripts to perform quality assessments and benchmarks. Evidence: Modules like modules/integration-testing.md and modules/performance-benchmarking.md implement logic that uses subprocess.run and os.access to call and verify the execution of local diagnostic tools.
[DYNAMIC_EXECUTION]: The evaluation workflow relies on the programmatic loading and execution of diagnostic scripts to generate metrics and performance data. Evidence: The scripts/README.md and modules/evaluation-workflows.md files describe a multi-phase discovery and analysis process that executes code from the scripts directory to evaluate skill efficiency and compliance.

skills-eval