The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process data from external sources (via WebSearch and specialized tools like agent-reach and qcc) which presents an inherent surface for indirect prompt injection. A malicious website could attempt to influence the analysis through embedded instructions.
Ingestion points: Phase 1 (Collection) retrieves findings from the open web, including social platform data and corporate registries.
Boundary markers: The skill proactively addresses this by requiring separate 'Collection' and 'Verification' agents, with the latter adopting an 'adversarial, default-skeptical' stance to identify misleading content.
Capability inventory: The system can spawn sub-agents (via the Workflow tool) and execute specific CLI tools (agent-reach, qcc) to verify data.
Sanitization: The skill employs a 'two-channel split' (FACTS vs COMMISSIONER_CONTEXT) which prevents sensitive user strategy or client names from being leaked into external search queries.
[DYNAMIC_EXECUTION]: The skill utilizes a JavaScript-based template (references/workflow_orchestration_template.md) to define the logic for the Workflow orchestration tool. This allows the agent to manage complex parallel agent pipelines at runtime.
Pattern: The orchestration script uses pipeline and agent functions to coordinate sub-tasks. This is a standard and supported method for complex agent workflows on the platform.
Context: The dynamic logic is limited to orchestration and follows a strict template provided within the skill's reference files.

benchmark-due-diligence