The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Python's subprocess module to execute the claude CLI tool and system utilities like lsof. This is found in scripts/run_eval.py, scripts/improve_description.py, and eval-viewer/generate_review.py. These operations are essential for running trigger evaluations and managing the lifecycle of the local evaluation viewer server.
[DATA_EXPOSURE]: The evaluation viewer script (eval-viewer/generate_review.py) reads files from the evaluation workspace and converts them to Base64 strings to display them in a local HTML report. This functionality allows developers to qualitatively review the output of their skills during the development process.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data such as user-provided test queries and feedback during its optimization loop.
Ingestion points: Reads data from eval_set.json, feedback.json, and the content of the skills being developed (e.g., SKILL.md).
Boundary markers: While the skill uses structured JSON for data exchange, it does not employ explicit 'ignore instructions' delimiters when interpolating data into prompts for the Claude CLI.
Capability inventory: The skill can execute shell commands via the claude CLI and perform file system operations within the local workspace.
Sanitization: Standard HTML escaping is applied to text data rendered in the browser-based evaluation viewer to prevent local script execution.

skill-creator