The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python scripts to automate complex workflows, including calling the claude CLI via subprocesses for benchmarking and triggering evaluation. It also uses lsof and os.kill in generate_review.py to manage local server ports, which is standard behavior for local development utilities.
[DATA_EXPOSURE]: The skill manages project-specific evaluation data and snapshots of existing skills within a local workspace directory. It adheres to standard secret management practices by referencing local file paths and does not perform any unauthorized data exfiltration.
[INDIRECT_PROMPT_INJECTION]: The skill has an inherent attack surface as it ingests user-provided test prompts (evals/evals.json) for evaluation. However, it relies on the safety filters of the subagents used to execute these prompts and provides explicit instructions to avoid creating malicious or misleading content.
Ingestion points: Test prompts are loaded from evals/evals.json and feedback.json.
Boundary markers: Execution occurs within separate subagent contexts which maintain their own safety boundaries.
Capability inventory: The skill can execute shell commands (subprocess), write files to the workspace, and start a local web server on 127.0.0.1.
Sanitization: The evaluation viewer uses secure rendering (e.g., textContent and XLSX.utils.sheet_to_html) to prevent cross-site scripting when displaying results.
[DYNAMIC_CONTEXT_INJECTION]: No !command syntax or silent execution of shell commands during skill load was detected in the SKILL.md frontmatter or body.

skill-creator