The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes "synthetic tools" that accept input from an agent to drive multi-step workflows, creating a surface where untrusted data could influence arguments passed to downstream tools.
Ingestion Points: User-provided data is ingested via the input_schema defined in synthetic tool configurations as described in references/synthetic-tools.md.
Boundary Markers: The documentation does not specify the use of explicit delimiters or instructions for the model to ignore embedded commands within the processed inputs.
Capability Inventory: Synthetic tools have the capability to execute sequences of API calls (e.g., create_ticket, search_contacts) across different connections.
Sanitization: The skill employs JSONata templates (args_template) for data interpolation but does not detail specific sanitization or escaping mechanisms for the input data before it is utilized in tool calls.
[DYNAMIC_EXECUTION]: The synthetic tool functionality uses JSONata expressions (args_template and output_template) to dynamically generate tool arguments and format results at runtime based on agent input and intermediate step results.
Evidence: The configuration schema in references/synthetic-tools.md defines args_template as a JSONata expression evaluated against { input, steps } to produce the argument object for each tool step.

elaichi-toolboxes