The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted meeting transcripts and summaries which could contain malicious instructions designed to manipulate the agent.
Ingestion points: Meeting transcripts, AI-generated summaries, and action items accessed via the meetings, recordings get-summary, and search subcommands in SKILL.md.
Boundary markers: There are no explicit delimiters or instructions specified to distinguish between system instructions and untrusted meeting content.
Capability inventory: The skill utilizes shell execution and supports writing output to local files or remote webhooks via the --deliver flag.
Sanitization: The skill documentation does not mention any sanitization or validation of the recording data before it is processed by the agent.
[DATA_EXFILTRATION]: The fathom-pp-cli tool includes a --deliver webhook:<url> flag that POSTs command results to a user-specified URL. This feature provides a direct mechanism for transferring sensitive meeting information to external endpoints.
[EXTERNAL_DOWNLOADS]: The skill downloads and installs binary tools from the vendor's infrastructure.
Details: Fetches the fathom-pp-cli and fathom-pp-mcp tools from the official github.com/mvanhorn/printing-press-library repository and installs the @mvanhorn/printing-press-library package via NPM.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute the fathom-pp-cli binary with various arguments, including subcommands like which and search that take natural language strings as input.

pp-fathom