The Agent Skills Directory

[SAFE]: The skill enforces a strict 'Human-in-the-loop' policy for all communications. It mandates that every reply must be explicitly approved by the user as 'exact-text' before the conversation.reply tool is executed, preventing the agent from sending unauthorized or malicious content.
[SAFE]: The skill addresses the attack surface of indirect prompt injection originating from inbound prospect replies.
Ingestion points: External data is ingested through the conversation.read operation in SKILL.md.
Boundary markers: The process defers to the approval-boundaries related skill to maintain safety constraints.
Capability inventory: The skill utilizes the conversation.reply (message writing) and conversation.classify (status updates) tools.
Sanitization: Instructions require the agent to minimize quotes from external communications and mandate human verification of all generated drafts before execution.
[SAFE]: The skill implements safety boundaries for permanent actions, such as recording 'Do-Not-Contact' dispositions, which require explicit human confirmation to prevent accidental or malicious blacklisting of contacts.
[SAFE]: The documentation explicitly identifies and warns against high-risk failure modes, including batch-sending drafts, silent channel switching, and loss of thread context, reinforcing a policy of transparency and manual oversight.

inbox-triage