The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the local repository (including README files, source code, and result outputs) to generate scientific prose. This creates a surface where instructions embedded in these files could potentially influence the agent's output. However, the skill implements strong grounding requirements to mitigate this risk.
Ingestion points: As defined in references/extraction.md, the skill identifies and reads files such as README*, docs/, *.md, configuration files, and result logs (CSV/JSON/TSV).
Boundary markers: The workflow requires the agent to cite evidence with path:line references and use specific [EVIDENCE] placeholders for missing metrics, which helps prevent the hallucination of data driven by injected prompts.
Capability inventory: The skill utilizes the agent's file-reading and text-generation capabilities. It suggests the use of a sandbox or processor to summarize data before bringing it into the context.
Sanitization: The skill emphasizes summarization and citation rather than raw interpolation of external content.
[EXTERNAL_DOWNLOADS]: The skill references official documentation and submission portals for major academic venues (e.g., SBC, ACM, IEEE). These are links to well-known academic services used for template verification and are considered safe.

to-article