The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes Python utility scripts (e.g., scripts/check_atlas_freshness.py) that execute git commands via subprocess.run. These commands are used to verify the repository state and fingerprint files included in the audit. The implementation correctly avoids shell=True and passes arguments as a list, which effectively mitigates command injection risks.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to audit external applications, meaning it ingests and processes untrusted data from source code, configuration files, and documentation. This ingestion represents a surface for indirect prompt injection.
Ingestion points: The agent reads the docs/product/application-atlas.md file, application source files defined in tracked_paths, and user interview responses.
Boundary markers: The skill employs a modular agent topology with explicit contracts for twelve distinct evaluator roles to isolate dimensions of analysis.
Capability inventory: The skill performs file system read/write operations for atlas management and evidence bundling, and executes git via subprocess.
Sanitization: Audit findings are validated against a strict schema using scripts/check_cognitive_overlay.py and the atlas integrity is checked by scripts/check_atlas_freshness.py before synthesis, reducing the risk of malicious payloads influencing the final report.

krt-product-polish-council