The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process large external corpora (text, code, or data), which presents an inherent surface for indirect prompt injection where malicious instructions in the data could attempt to influence subagent outputs.
Ingestion points: Worker subagents read raw corpus segments as their primary input in SKILL.md Step 5.
Boundary markers: The skill incorporates strong defensive prompting by explicitly instructing subagents: "Do not dispatch or invoke any subagents," which mitigates the risk of recursive execution attacks.
Capability inventory: The orchestrator utilizes TaskCreate and Bash tools; however, subagents are strictly limited to reading their assigned inputs and writing reports to specific local paths.
Sanitization: No specific content sanitization is mentioned, but the workflow is designed to treat inputs as passive data for synthesis.
[DYNAMIC_EXECUTION]: The skill uses shell-based Python execution for orchestration logic and layout planning.
Evidence: The skill uses python3 -c to perform arithmetic and executes the bundled compute_layout.py script to calculate token counts and agent assignments. These are legitimate, local-only operations for the skill's primary purpose.
[COMMAND_EXECUTION]: Standard Bash tools are used for file system management and data estimation.
Evidence: The skill utilizes wc -c, find, mktemp, and mkdir to prepare the working directory and calculate the layout based on file sizes.

doing-a-simple-two-stage-fanout