The Agent Skills Directory

[PROMPT_INJECTION]: Use of imperative and overriding language to enforce specific workflows and tool behavior. Evidence: The skill employs markers like 'CRITICAL' and 'MUST execute' to force adherence to specific sub-routines (e.g., the AskUserQuestion guard and mandatory artifact persistence) regardless of user input or agent state.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from web sources and processes it using powerful tools without sanitization or boundary markers. Ingestion points: Data enters the context through WebSearch and WebFetch tools used during the research phase, as well as the $ARGUMENTS variable. Boundary markers: Absent. There are no instructions to delimit or ignore potentially malicious instructions embedded in the external content fetched from the web. Capability inventory: The skill has access to Write (local file system), Task (spawning sub-agents), and Read/Glob/Grep (file system discovery and reading). Sanitization: Absent. External content is processed and merged into the decision-making matrix and thinking artifacts without validation.

investigate