The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits a surface for indirect prompt injection (Category 8) by processing untrusted data from the local environment.
Ingestion points: The skill reads research briefs and source lists from the literature-reviews/ directory based on a user-provided slug.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to treat the content of the research briefs as data rather than instructions.
Capability inventory: The skill is capable of network operations via WebFetch/WebSearch and file system writes through the ane_package.reporting module.
Sanitization: While a "prose-fidelity gate" using a secondary model is mentioned, this is an LLM-based check rather than a strict security sanitization process.
[EXTERNAL_DOWNLOADS]: The skill performs automated network requests using WebFetch and WebSearch to audit and canonicalize URLs found within the research documents. This could be triggered by malicious URLs embedded in the source files.
[COMMAND_EXECUTION]: The skill executes methods from the internal ane_package.reporting library (e.g., write_pptx_deck, write_word_report) to generate and save output files to the local file system.

learning-product