The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple scripts, including scripts/run_eval.py and scripts/improve_description.py, utilize subprocess to execute the claude CLI and system utilities like lsof. These calls are restricted to the local environment and are used to automate the testing and iteration of skill drafts.\n- [EXTERNAL_DOWNLOADS]: The evaluation viewer (eval-viewer/viewer.html) loads the SheetJS library from cdn.sheetjs.com. This is a trusted service used for local client-side processing of spreadsheet files.\n- [DATA_EXFILTRATION]: eval-viewer/generate_review.py hosts a local HTTP server on 127.0.0.1 to display evaluation results. The server transmits workspace data, including output files from test runs, to the user's browser. While this involves local data transmission, the server is bound to the localhost interface.\n- [PROMPT_INJECTION]: The skill provides an indirect prompt injection surface. Ingestion points: User-defined prompts in evals/evals.json. Boundary markers: Absent. Capability inventory: subprocess calls in scripts/ and local file hosting in eval-viewer/. Sanitization: Absent. This surface is functional for testing agent behavior under various scenarios.\n- [SAFE]: Analysis of the source code confirms that its high-privilege operations are aligned with its purpose as a development tool. No persistence mechanisms, credential harvesting, or unauthorized remote access patterns were found.

skill-creator