The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill takes user input via the QUESTION argument and can ingest local files through the --context-file flag. This content is passed to multiple LLMs for synthesis, creating a surface for indirect prompt injection where adversarial instructions in the data could influence the final verdict.
Ingestion points: QUESTION command-line argument and content from files specified by the --context-file argument.
Boundary markers: No specific delimiters or safety instructions are defined in the skill text to separate user data from the system instructions.
Capability inventory: The skill executes the uv run researchkit command, which involves network communication with LLM providers and file system reads.
Sanitization: No explicit sanitization or validation of the input data is described.
[COMMAND_EXECUTION]: The skill relies on the execution of uv run researchkit in the shell. While this is the primary intended mechanism of the skill, it assumes that the local researchkit repository is trusted and that the uv tool is installed in the environment.

council