The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data, including firing alerts, incident postmortems, and user-provided outage descriptions, to generate remediation commands. This creates a surface where malicious instructions embedded in an alert payload or postmortem could influence the generated runbook.
Ingestion points: Alert text, alert payloads, dashboard panel descriptions, and postmortem documents (referenced in SKILL.md).
Boundary markers: The skill uses a structured Markdown template (RUNBOOK_TEMPLATE.md) for output but does not define explicit delimiters or 'ignore' instructions for the input data.
Capability inventory: The agent generates shell commands for diagnosis, remediation, and rollback. While it does not execute these commands autonomously (requiring explicit human approval), the generation of incorrect or harmful commands is a primary risk.
Sanitization: The skill includes guardrails instructing the agent to avoid hardcoded secrets and to flag unverified steps, but it lacks automated sanitization for input data.
[COMMAND_EXECUTION]: The skill is designed to generate executable shell commands for operational recovery. It mitigates risk by requiring that commands be idempotent, read-only during diagnosis, and reviewed by a human before execution in production environments.

runbook-authoring