The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external SKILL.md files which represent an ingestion point for untrusted data. Instructions embedded in audited skills could attempt to influence the agent's behavior during the review process.
Ingestion points: The workflow in SKILL.md uses repo.rglob('SKILL.md') to discover and read external instruction files for analysis.
Boundary markers: The audit methodology does not explicitly define delimiters or 'ignore instructions' warnings when reading audited content.
Capability inventory: The skill utilizes execute_code (Python) and file read/write operations across its scripts.
Sanitization: Analysis is performed via regular expressions and YAML parsing; however, the agent interprets the final audit report.
[DATA_EXFILTRATION]: The skill accesses the agent's internal state database to generate usage statistics. While the data is used locally for analytics, it involves reading sensitive metadata files.
Evidence: SKILL.md contains a Python script that connects to the agent's local database using sqlite3.connect(str(Path.home() / '.hermes' / 'state.db')) to count tool invocations.
[COMMAND_EXECUTION]: The skill encourages the use of shell-like operations and programmatic file system access to perform batch audits.
Evidence: The skill uses pathlib and os.walk to traverse directories and perform regex-based analysis on files within a repository.
[DYNAMIC_EXECUTION]: The skill facilitates the generation and execution of Python code snippets to perform quality checks.
Evidence: references/batch-checks.md provides multiple Python code blocks intended to be run via the execute_code tool to validate frontmatter, check for identity leakage, and calculate trigger ratios.

skill-quality-review