The Agent Skills Directory

[PROMPT_INJECTION]: The instructions and metadata are focused on the skill's primary purpose. No attempts to override agent behavior or bypass safety guidelines were found.
[CREDENTIALS_UNSAFE]: No hardcoded secrets were detected. The configuration template (assets/configuration_template.yaml) correctly uses placeholders like 'YOUR_DATABASE_PASSWORD' for user-provided credentials.
[EXTERNAL_DOWNLOADS]: The skill does not perform any remote downloads of scripts or data from untrusted sources.
[DATA_EXFILTRATION]: There are no network operations or access requests for sensitive system files (e.g., .ssh, .env). The scripts are limited to local file system operations.
[INDIRECT_PROMPT_INJECTION]: The skill defines an ingestion surface for untrusted data (user-item interactions).
Ingestion points: Data files processed via scripts/data_preprocessing.py and other utility scripts.
Boundary markers: No explicit delimiters are used to separate untrusted data from the agent context.
Capability inventory: The skill has access to powerful tools including Bash(cmd:*) and Write.
Sanitization: No sanitization of input data is performed. However, as the provided scripts only perform basic file metadata extraction and JSON validation, the risk of instruction execution from data is low.

building-recommendation-systems