The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Python subprocess module to execute system commands. In scripts/run_eval.py and scripts/run_loop.py, it invokes the claude CLI to run evaluations of other skills. In eval-viewer/generate_review.py, it executes lsof to manage network ports used by the evaluation viewer's local HTTP server.
[EXTERNAL_DOWNLOADS]: The eval-viewer/viewer.html file includes a script tag that loads the SheetJS library from cdn.sheetjs.com. This download is from a well-known service and is used for rendering spreadsheet data within the evaluation viewer.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface (Category 8) because it is designed to ingest and process untrusted external content as part of the skill creation and optimization workflow.
Ingestion points: Untrusted data enters the agent context via evals/evals.json, feedback.json, and user-provided skill drafts.
Boundary markers: The skill uses XML-style tags such as <new_description> to delimit generated content, though these do not provide a complete defense against adversarial content.
Capability inventory: The skill has capabilities including filesystem write access, network communication via the Anthropic API, and subprocess command execution.
Sanitization: The evaluation viewer uses basic HTML escaping to prevent XSS during result rendering, but the core logic relies on prompt instructions rather than robust data sanitization for instruction isolation.

skill-creator