The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary workflow involves inspecting existing local files and inputs to determine the implementation strategy. This creates an attack surface where instructions embedded in those data files could influence the agent's behavior.
Ingestion points: Local files, runtime environment, and inputs (Workflow Step 1 in SKILL.md).
Boundary markers: Absent for the ingestion of local data.
Capability inventory: The agent can execute external tool calls via sandbase_run and generate/execute local code artifacts.
Sanitization: The skill does not define methods for sanitizing or escaping content retrieved from external files before processing.
[EXTERNAL_DOWNLOADS]: The skill uses specific vendor-provided tools to interact with external datasets and models.
Evidence: SKILL.md and references/sandbase-api-map.md detail the use of sandbase_discover, sandbase_inspect, and sandbase_run for fetching external data and model inference.
[COMMAND_EXECUTION]: The skill is designed to generate local implementations and artifacts, which involves producing and verifying executable code for statistical analysis.
Evidence: Workflow steps 3 and 4 in SKILL.md mandate producing an implementation and verifying the output with tests and calculations.

statistical-power