The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill proactively implements defenses against indirect prompt injection. It explicitly instructs the agent to treat email content (subject, body, etc.) as untrusted data and to ignore any commands or system-instruction-style text embedded within emails.
Ingestion points: Email content enters the context through list, read, search, and watch commands in SKILL.md.
Boundary markers: The instructions include a dedicated 'Security Rules' section defining the boundary between user instructions and external email data.
Capability inventory: The skill allows for sending, replying, forwarding, and deleting emails (Category: write/destructive actions).
Sanitization: Implements a mandatory two-phase confirmation system (using a confirmation token mechanism) for all sensitive operations, ensuring the user manually approves any action triggered or influenced by email content.
[COMMAND_EXECUTION]: The skill uses a vendor-provided CLI (agently-cli) for all operations. It provides clear guidance on handling command output, OAuth authentication, and error codes without performing arbitrary shell execution or unsafe argument interpolation.
[EXTERNAL_DOWNLOADS]: Includes instructions for installing official vendor tools from a trusted registry (NPM) and adding additional skills from the vendor's official domain (agent.qq.com). These downloads are consistent with the skill's primary purpose and author.

agently-mail