The Agent Skills Directory

[SAFE]: The skill emphasizes human oversight by requiring the agent to pause and request user confirmation before performing irreversible or sensitive actions, such as spending money, deleting files, or communicating with external systems.\n- [SAFE]: In technical environments like Codex, the skill utilizes the propose_scope tool to define and restrict the agent's file system access to specific allowed paths, minimizing the risk of unauthorized file modifications.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes external data through the ask_user tool when using the input type, which presents a potential surface for indirect instructions if the input source is compromised.\n
Ingestion points: The value returned by the ask_user tool (SKILL.md).\n
Boundary markers: The instructions do not define specific delimiters or guardrail prompts to be used when the agent processes strings returned from the external interface.\n
Capability inventory: The agent can trigger notifications, poll for answers, and, in Codex environments, execute shell commands and modify files.\n
Sanitization: There are no explicit instructions requiring the agent to sanitize or validate the content received from the user's mobile device or dashboard before acting upon it.

pushary-chatgpt