The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted external data during its OSINT verification and collection stages.
Ingestion points: External data enters the agent context via browser automation tools specified in Stage 3 (Collection) and Stage 4 (Verification) of SKILL.md and subagents/.cursor/agents/verification-expert.md.
Boundary markers: The methodology mandates a falsify-first approach and includes an explicit references/ETHICS_GUARDRAILS.md file that provides instructions on refusing malicious or problematic requests embedded in external data.
Capability inventory: The skill uses browser automation (screenshots, snapshots) and subagents for hypothesis testing and report generation across multiple files.
Sanitization: The skill incorporates a robust verification process including a dedicated qa-gatekeeper.md subagent and a references/QA_CHECKLIST.md to fact-check findings and verify source credibility before delivery.
[METADATA_POISONING]: Several files in the subagents directory contain significant metadata anomalies.
Evidence: subagents/.cursor/agents/README.md contains over 50 identical copies of its YAML frontmatter. subagents/.cursor/agents/research-lead.md contains three redundant copies of its frontmatter. While not containing malicious instructions, this represents severe metadata redundancy that may unnecessarily consume context tokens or attempt to influence model focus through repetition.

deepresearch