The Agent Skills Directory

[METADATA_POISONING]: The skill description uses the prefix 'CRITICAL:' followed by a list of trigger keywords. This formatting is designed to override standard agent priority and force the selection of this skill by mimicking system-level alerts. Additionally, the frontmatter explicitly sets 'risk: critical', which is a deceptive metadata field intended to influence the agent's perception of the skill's importance.
[INDIRECT_PROMPT_INJECTION]: The skill implements logic for parsing and displaying data from external sources, specifically SSE (Server-Sent Events) streams from LLM providers in the 'parse_sse' function and 'BotClient' trait. Ingestion points: External data enters the context via the 'parse_sse' stream and is stored in 'MessageContent' within SKILL.md. Boundary markers: The instructions do not define delimiters or specific 'ignore' instructions for the content being processed. Capability inventory: The skill uses 'spawn()' for async task execution, performs network requests via 'client.post()', and renders dynamic content in UI 'Slot' widgets. Sanitization: There is no explicit evidence of sanitization or validation of the text delta received from the LLM stream before it is appended to the message content.

molykit