The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocesses to execute the claude CLI tool (scripts/run_eval.py) for testing skill triggering and manages local server processes using lsof and kill (eval-viewer/generate_review.py). These operations are necessary for the tool's core functionality of benchmarking and serving the evaluation viewer. \n- [EXTERNAL_DOWNLOADS]: The skill communicates with the Anthropic API to perform analysis and improve skill descriptions (scripts/improve_description.py). Additionally, the evaluation viewer template (eval-viewer/viewer.html) loads the SheetJS library from a well-known CDN for rendering spreadsheet data. All external references are to trusted organizations or established technology service providers. \n- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it ingests untrusted data from evaluation sets and user feedback which are then processed by the agent. \n
Ingestion points: Evaluation queries are read from eval_set.json in scripts/run_eval.py, and user reviews are read from feedback.json in eval-viewer/generate_review.py. \n
Boundary markers: The skill uses XML-style tags such as <skill_content> and <attempt> in scripts/improve_description.py to delimit untrusted data within LLM prompts. \n
Capability inventory: Across its bundled scripts, the skill can execute the claude CLI, start a local HTTP server, and write files to the local workspace. \n
Sanitization: The skill employs standard JSON parsing and YAML block scalars to prevent structural breakages, relying on model reasoning for data interpretation.

skill-creator