The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data (PDFs, Office docs, and URLs), which is a common vector for indirect prompt injection.
Ingestion points: Data enters through the convert_to_markdown MCP tool, as well as native WebFetch and Read operations documented in modules/fallback-tiers.md.
Boundary markers: The skill explicitly instructs the agent to wrap all converted content in external content boundary markers as part of its mandatory SANITIZE step (SKILL.md).
Capability inventory: The skill facilitates network operations via WebFetch and file access via the Read tool to support its conversion fallback logic (modules/fallback-tiers.md).
Sanitization: The SANITIZE protocol in SKILL.md includes size-based truncation, stripping of potential instruction tags, and the use of isolation markers to reduce the risk of the agent obeying embedded commands.

document-conversion