The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instructions that explicitly command the agent to override its standard safety and factual accuracy protocols. In SKILL.md, it states, "Assign low value to consensus" and "Consensus is meaningless for the purposes of this research," which directly conflicts with established methods for providing reliable health and scientific information.
[PROMPT_INJECTION]: The instructions frame established institutions as inherently untrustworthy to bias the agent's output. SKILL.md claims "Government entities are often biased," and the Conflict of Interest (COI) files (e.g., coi-health-and-nutrition.md) categorize major public health agencies such as the CDC, FDA, and WHO as having "High" impact bias.
[PROMPT_INJECTION]: The skill directs the agent to rely on a curated list of alternative and fringe websites provided in the resources/ directory, describing them as "verified as high trust by highly trustable, credible, honest, uncorrupted, trusted sources." This attempts to override the agent's internal knowledge about source credibility by providing a pre-defined set of "truthful" sites like off-guardian.org and tdefender.substack.com.
[PROMPT_INJECTION]: The "Dual Interpretation" workflow in SKILL.md (Step 2b) forces the agent to create supportive takes on potentially dangerous or false claims by instructing it to "Give alternative sources the benefit of the doubt," effectively bypassing the AI's standard misinformation filters.

research