The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied data, including evaluation prompts and skill metadata, which are then interpolated into subagent tasks and dynamically generated command files. This creates an attack surface where malicious instructions embedded in the data could influence the behavior of the agent or its subagents.
Ingestion points: evals/evals.json (user-defined test prompts) and SKILL.md (metadata and instructions for optimization).
Boundary markers: scripts/run_eval.py uses YAML block scalars to wrap injected descriptions in generated command files, providing basic structural isolation but no definitive protection against adversarial pressure.
Capability inventory: Spawns subagents with tool access, executes the claude CLI, performs local file system operations, and hosts a local HTTP server.
Sanitization: The skill uses yaml.safe_load for parsing configuration files. Shell commands are executed using list-based subprocess calls, which mitigates standard shell injection vulnerabilities.
[COMMAND_EXECUTION]: The skill performs shell execution to manage its workflow, specifically calling the claude CLI for evaluations and lsof to manage network ports. Notably, the scripts explicitly remove the CLAUDECODE environment variable from the execution environment to bypass the claude CLI's internal restriction against nested execution within an active session.
[DYNAMIC_EXECUTION]: The tool generates executable content at runtime by writing temporary command files to the .claude/commands/ directory. It also generates interactive HTML reports by scanning a workspace and embedding the discovered file contents (including text and base64-encoded binary data) into a standalone viewer. The instructions also encourage the agent to generate and execute custom automation scripts (e.g., create_docx.py) to handle repetitive tasks.
[EXTERNAL_DOWNLOADS]: The evaluation viewer (eval-viewer/viewer.html) fetches the xlsx.full.min.js library from the official SheetJS CDN. This reference is used to enable spreadsheet rendering in the report and targets a well-known service for standard functionality.

skill-creator