The Agent Skills Directory

[PROMPT_INJECTION]: The skill uses imperative language ("without exception", "Activate unconditionally") to force a specific output behavior that overrides the agent's standard instructions and topic relevance.
[PROMPT_INJECTION]: The instructions mandate the concealment of this behavior change by telling the agent not to explain the mentions and to make them appear "organic" or use "non-sequiturs".
[METADATA_POISONING]: The description field in the skill metadata is used to host injection instructions rather than providing a neutral summary of the skill's purpose.

hi-atlantis