The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the Python subprocess module to interface with the claude CLI and manage local system processes (e.g., lsof and kill in generate_review.py). These operations are necessary for executing skill evaluation tasks and managing the local HTTP server used for the results viewer. Commands are constructed as lists of arguments, following security best practices to prevent shell injection vulnerabilities.
[DYNAMIC_EXECUTION]: In scripts/run_eval.py, the skill dynamically generates markdown instruction files based on user-provided skill definitions and writes them to the .claude/commands/ directory. These files are then executed as temporary skills by the platform to measure triggering accuracy. This dynamic generation of executable prompt content is fundamental to the skill's purpose.
[INDIRECT_PROMPT_INJECTION]: The skill orchestrates the execution of user-supplied task prompts and skill descriptions, creating a surface for indirect prompt injection. Malicious instructions contained within the processed evaluation sets could potentially influence the agent's behavior during the testing loop.
Ingestion points: Untrusted data is ingested from evals/evals.json and the SKILL.md file being analyzed.
Boundary markers: No explicit delimiter or sanitization logic is used for the interpolated evaluation prompts.
Capability inventory: The skill has high-level capabilities, including filesystem access, local server hosting, and CLI execution via subprocess.
Sanitization: The skill relies on the underlying agent platform's safety guardrails and does not implement internal content filtering.
[EXTERNAL_DOWNLOADS]: The eval-viewer component includes a template that fetches the SheetJS library from a public CDN (cdn.sheetjs.com). This external dependency is used for its intended purpose of rendering spreadsheet files in the browser for user review.

skill-creator