The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation in SKILL.md provides installation instructions that download a CLI tool from npm (@mvanhorn/printing-press-library) and GitHub (mvanhorn/printing-press-library). These sources are owned by the skill author and represent the intended installation path.
[DATA_EXFILTRATION]: The skill features a delivery mechanism (--deliver webhook:) in SKILL.md that can POST command output to arbitrary external URLs. Additionally, the auth login command stores session tokens locally at ~/.superhuman-pp-cli/ for persistent authentication as described in the documentation.
[PROMPT_INJECTION]: The skill facilitates the ingestion of untrusted email data while possessing high-privilege capabilities, creating an indirect injection surface. 1. Ingestion points: SKILL.md includes commands for reading emails such as threads list, messages list, and semantic search (ai). 2. Boundary markers: Boundary markers are absent; the skill does not define specific delimiters to separate untrusted email content from agent instructions. 3. Capability inventory: SKILL.md documents capabilities to send emails, create drafts, and write raw CRDT user data. 4. Sanitization: Sanitization is absent; no filtering or validation of email content is described.
[COMMAND_EXECUTION]: The userdata write command in SKILL.md provides a low-level escape hatch to write raw JSON values to Superhuman CRDT paths, which could be misused to modify account state if an agent is influenced by malicious input from ingested emails.

pp-superhuman