The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data through orchestrated workflows and includes built-in mitigations. Ingestion points include file scouting and node-specific scopes; boundary markers are established via the node-brief template which instructs agents to ignore embedded instructions. The capability inventory includes the Workflow and Skill tools, while sanitization is achieved through schema-enforced JSON outputs and an adversarial verification phase that refutes findings against original source code.\n- [DYNAMIC_EXECUTION]: The skill documents the use of the platform's Workflow tool, which utilizes JavaScript for orchestration. The provided templates follow safety guidelines by avoiding non-deterministic functions and using structured data passing to manage sub-agent lifecycle and state.\n- [COMMAND_EXECUTION]: The architecture separates read-only analysis from write operations. It mandates that sub-agents for review tasks have no write access and requires explicit user approval and frontier-based scheduling for any graph that involves modifying files.

graph-engineering