The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python's subprocess and os modules across several utility scripts (run_eval.py, run_loop.py, generate_review.py) to execute the claude CLI and local system commands (e.g., lsof for port management). This execution is the intended mechanism for running skill evaluations and managing the local viewer server.
[EXTERNAL_DOWNLOADS]: The README.md documentation provides manual installation steps that involve cloning a repository from GitHub (github.com/alenazaharovaux/share). This is documented as the vendor's own repository for distributing the skill content.
[PROMPT_INJECTION]: The skill contains an indirect prompt injection surface (Category 8) within its description optimization loop (scripts/run_loop.py). The tool ingests user-provided or agent-generated evaluation queries from evals/evals.json and interpolates them into a prompt used by a secondary LLM to refine the skill's description.
Ingestion points: evals/evals.json, feedback.json, and direct user input for evaluation queries.
Boundary markers: The prompt in improve_description.py uses XML-style tags (<current_description>, <scores_summary>, <skill_content>) to delimit untrusted data.
Capability inventory: The skill possesses file-write access, network access (via the Anthropic API client), and local command execution via subprocess.
Sanitization: While inputs are structured via JSON, the core natural language content of queries is processed by the LLM without specific sanitization against adversarial instructions.

skill-creator