The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill documents the capability for the agent to author and run JavaScript scripts at runtime to orchestrate tasks. This includes passing scripts inline or loading them from session-persisted file paths.\n
Evidence: The reference describes passing scripts via the script parameter, re-invoking workflows using {scriptPath: "<path>"}, and executing code in an async JavaScript context.\n- [INDIRECT_PROMPT_INJECTION]: The orchestration framework processes data through a pipeline of subagents, creating a vulnerability surface for indirect prompt injection if data is not properly isolated or sanitized.\n
Ingestion points: Untrusted data enters the system through the args global or as output from previous agent() calls.\n
Boundary markers: The documentation does not specify the use of delimiters or 'ignore' instructions for data passed to subagents.\n
Capability inventory: Scripts can spawn subagents (agent()), execute nested workflows (workflow()), and access tools via ToolSearch.\n
Sanitization: No explicit guidance is provided for escaping or validating external content before it is interpolated into agent prompts.\n- [PROMPT_INJECTION]: The skill contains instructions that encourage the agent to override default resource constraints under specific conditions.\n
Evidence: The 'Ultracode' section instructs the agent that 'token cost is not a constraint' and to 'author and run a workflow for every substantive task by default' when a specific system reminder is present.

workflow-authoring