The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run and subprocess.Popen in several scripts, including scripts/run_eval.py, scripts/improve_description.py, and scripts/run_loop.py. These calls are used to invoke the claude CLI for automated testing and optimization of skill descriptions.
[DATA_EXFILTRATION]: The eval-viewer/generate_review.py script facilitates result review by starting a local HTTP server bound to 127.0.0.1. It reads content from the local workspace to render an interactive HTML report.
[DYNAMIC_EXECUTION]: The skill programmatically generates temporary skill configuration files within the .claude/commands/ directory during evaluation iterations to test triggering accuracy.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection as it ingests and processes external data to guide agent behavior.
Ingestion points: Untrusted data enters the context via evals/evals.json (user-defined test prompts) and feedback.json (human review comments).
Boundary markers: The skill utilizes XML-style tags (e.g., <new_description>) to structure its internal prompting, but lacks robust delimiters to explicitly prevent the execution of instructions embedded within the test queries themselves.
Capability inventory: The skill has permission to execute shell commands (via the claude CLI) and perform read/write operations on the local file system.
Sanitization: The scripts/generate_report.py script employs html.escape when building HTML reports to mitigate basic injection risks in the viewer interface.

create-skill