The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill provides templates for multi-agent orchestration that are susceptible to indirect prompt injection due to the lack of input sanitization and explicit boundary markers in the demonstrated prompt constructions.
Ingestion points: Untrusted data enters the workflow via initial_input in the SequentialPipeline.run method, user_request in the HierarchicalOrchestrator.run method, and competitors list in parallel_research.
Boundary markers: The code snippets interpolate data directly into prompts (e.g., messages=[{"role": "user", "content": stage_input}]) without using delimiters or instructions to ignore embedded commands.
Capability inventory: The skill's primary capability involves network operations using the anthropic Python library to interact with external LLM APIs.
Sanitization: There is no evidence of sanitization or validation for external content. The HierarchicalOrchestrator class is specifically vulnerable to schema confusion, as it uses json.loads to parse and act upon a task plan generated by an LLM without verifying the structure or content of the resulting JSON object.

agent-workflow-designer