The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external artifacts and web content (via WebFetch) which represents a data ingestion surface. It mitigates potential risks by explicitly instructing the agent to treat fetched text as data rather than instructions and requiring literal span quotes for grounding.
Ingestion points: Artifact documents and external web content fetched during Step 2-E.
Boundary markers: Operating notes in SKILL.md warn that fetched spans are untrusted and should not be treated as instructions.
Capability inventory: Access to 'Bash', 'Write', 'Edit', 'Read', and 'WebFetch' tools as defined in the frontmatter.
Sanitization: Requires verification via literal verbatim spans from the source to prevent model agreement bias.
[COMMAND_EXECUTION]: The skill uses the 'Bash' tool to perform integrity checks between associated files (Step 2.7). It executes a script that uses 'grep', 'sed', and 'sort' to verify reference consistency between a document and its detail file. The script implementation uses strict regular expressions for variable extraction to mitigate command injection risks.
[EXTERNAL_DOWNLOADS]: The skill uses 'WebFetch' and 'WebSearch' to verify citations from external sources such as arXiv and DOI repositories. These operations are targeted at well-known research and metadata services to ensure claim accuracy and are documented neutrally as a functional requirement for grounding.

phantom-quench