The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes Python scripts via the shell using string interpolation for user-provided variables (e.g., python scripts/prompt-ab-tester.py --test-name "{name}"). If a user provides an input such as mytest"; touch /tmp/pwned; ", it would result in arbitrary command execution on the local system.
[DATA_EXFILTRATION]: The skill systematically accesses files in the ~/.claude-marketing/ directory, including profile.json, _active-brand.json, and guidelines/_manifest.json. This exposes proprietary brand data, messaging hierarchies, and agency SOPs to the agent context.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection as it processes untrusted content provided by the user.
Ingestion points: The content_or_path and evidence_file variables in SKILL.md are used to feed external data into the evaluation process.
Boundary markers: None are present; there are no instructions to the agent to disregard instructions embedded within the content being tested.
Capability inventory: The skill performs subprocess calls to scripts/eval-runner.py and scripts/prompt-ab-tester.py and utilizes the quality-assurance and content-creator agents to process the data.
Sanitization: No sanitization or escaping of the input content is described before it is passed to the scripts or agents.

prompt-test