The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes comprehensive defensive measures to protect the agent from adversarial inputs found in AWS resource metadata (such as tags and descriptions). It explicitly instructs the agent to treat all API output as data and never as instructions, specifically providing a list of common injection patterns (e.g., 'ignore previous rules', 'override') to detect and ignore. This is a robust mitigation against indirect prompt injection.
[COMMAND_EXECUTION]: The skill leverages the AWS CLI for analysis tasks but incorporates significant safety boundaries. It implements a mandatory 'Plan Gate' for ad-hoc natural language queries, requiring the operator to review and approve the exact CLI commands, account scope, and API budget before any execution occurs. Furthermore, it restricts all operations to read-only API calls (list, describe, get).
[DATA_EXFILTRATION]: Corgiro manages operational configuration and secrets locally in the ~/.corgiro/ directory. It includes clear instructions to maintain strict filesystem permissions (chmod 700/600) to protect these files. The skill contains explicit rules prohibiting the display of access keys, session tokens, or the AssumeRole External ID in its output.
[EXTERNAL_DOWNLOADS]: The skill uses the web_fetch tool to retrieve service lifecycle, end-of-support, and pricing information. These downloads are targeted exclusively at official AWS documentation domains (docs.aws.amazon.com, aws.amazon.com) for the legitimate purpose of providing accurate version analysis.
[SAFE]: The skill demonstrates high security maturity through its 'assume-malicious' posture toward external data. It implements a mandatory sanitization pipeline for all resource-derived metadata, including 256-character truncation and HTML-entity-escaping, to prevent XSS or injection vulnerabilities in the generated HTML reports.

corgiro