The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from the web and local file systems to generate presentation slides, creating a vulnerability surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through moda web search and moda web read commands, as well as via local file-reading and search tools used during the gathering phase.
Boundary markers: The skill does not explicitly provide boundary markers or instructions to ignore embedded directives within the ingested data, although it emphasizes verbatim data preservation.
Capability inventory: The agent has the ability to execute numerous moda CLI commands (authorized via Bash(moda:*)) to create and modify canvas content, generate AI media, and export files.
Sanitization: The skill relies on a markup parser that strips disallowed HTML tags and a sandboxed JavaScript environment (moda canvas edit) that blocks sensitive operations like fetch, eval, and require.
[EXTERNAL_DOWNLOADS]: The skill downloads brand assets from temporary signed URLs for visual inspection as part of the brand verification workflow.
Evidence: references/brand.md contains instructions to download files using curl -o /tmp/brand-logo-1.png "<url>".
[COMMAND_EXECUTION]: The skill is configured to execute commands exclusively through the moda toolset.
Evidence: The SKILL.md file defines allowed-tools as Bash(moda:*).
[DYNAMIC_EXECUTION]: The skill utilizes a sandboxed JavaScript environment to perform batch mutations and logic-heavy edits on the presentation canvas.
Evidence: references/edit-code.md describes the moda canvas edit command which runs synchronous JavaScript to update nodes, pages, and variables.

moda-deck