The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the claude CLI via subprocess in scripts/run_eval.py and scripts/improve_description.py to test triggering accuracy and optimize skill descriptions. In scripts/run_eval.py, the code explicitly removes the CLAUDECODE environment variable before execution to bypass internal guards that prevent nested CLI sessions.
[DYNAMIC_EXECUTION]: The skill is designed to generate, patch, and package executable content, including Python scripts (e.g., scripts/init_skill.py) and markdown instructions. These are then intended to be installed and executed as independent skills, creating a recursive execution capability.
[INDIRECT_PROMPT_INJECTION]: The skill implements an iteration loop that ingests user-provided intent and prompts to generate and refine instructions for AI agents. This creates a significant surface for indirect injection, as malicious user input could influence the behavior of the generated skills. The framework relies on manual human review of feedback in the viewer to mitigate these risks.
[PRIVILEGE_ESCALATION]: When operating in specific sandboxed environments (DeerFlow), the skill instructs the agent to use the skill_manage tool to bypass standard sandbox filesystem isolation and persist modifications directly to the user's persistent skill directory.
[EXTERNAL_DOWNLOADS]: The eval-viewer/viewer.html file includes a reference to the xlsx library hosted on cdn.sheetjs.com to enable the rendering of spreadsheet data within the evaluation review interface.

skill-creator