The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves fetching and summarizing external academic papers, which creates a surface for indirect prompt injection attacks. Malicious instructions embedded in a paper's text could attempt to hijack the agent's behavior during synthesis.
Ingestion points: Untrusted content from papers is extracted and stored in source.md and card.md files under the research/collection/ directories.
Boundary markers: The skill uses a structured metadata schema (card.md) to manage paper data, which provides some structural isolation, but the primary text remains unverified.
Capability inventory: The agent can write to the file system, manage git issues and branches via project:epic-dev, and execute CLI commands via uvx.
Sanitization: There are no explicit instructions to sanitize the text extracted from external sources for adversarial prompt patterns.
[EXTERNAL_DOWNLOADS]: The skill depends on the opencite CLI utility, invoked via the uvx package runner, to retrieve paper metadata and PDFs. This is a functional dependency appropriate for the skill's intended use in academic research.

lit-review