The Agent Skills Directory

[PROMPT_INJECTION]: The skill is designed to perform internet-backed research, which exposes the agent to indirect prompt injection. Untrusted data from the web is ingested and processed, potentially allowing external instructions to influence the agent's reasoning or output.
Ingestion points: The workflow in SKILL.md and references/web_research_protocol.md requires the agent to conduct broad and focused internet searches, fetching content from at least 8 to 12 different sources.
Boundary markers: The skill uses structured intermediate files like source_dossier.md and claim_evidence_ledger.md to organize and isolate evidence, though it does not employ explicit character-based delimiters for raw external text.
Capability inventory: The agent possesses capabilities to perform network searches and generate structured markdown files. It interacts with other skills (e.g., infographic-powerpoint-deck) via the slide extraction table.
Sanitization: The risk is mitigated by an extensive set of QA gates, including references/logic_qa_checklist.md, references/evidence_gates.md, and references/objectivity_checks.md. These checks enforce logical consistency, evidence triangulation, and neutrality, making it difficult for malicious instructions from a single source to compromise the final article.

deep-research-article