The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates communication between a human and an agent, where human-provided responses are returned to the agent's context. This creates a surface for potential instruction injection from the human user. In this context, it is the intended primary functionality designed to provide human oversight (human-in-the-loop) for agent actions.\n
Ingestion points: Human input is captured through the ask, approve, select, and approve-steps commands documented in SKILL.md and references/cli-reference.md.\n
Boundary markers: The tool uses structured JSON output formats and specific exit codes to distinguish between approved actions, answers, and errors.\n
Capability inventory: Documentation examples include using the human's response to conditionally execute file deletions (rm -rf), build commands (npm run build), or deployment scripts (kubectl apply), as seen in SKILL.md and references/cli-reference.md.\n
Sanitization: The protocol relies on typed 'pings' and structured data schemas to isolate user content from the agent's control logic.

agentping