The Agent Skills Directory

[COMMAND_EXECUTION]: The skill workflow requires executing several local Python scripts via the shell to perform quantitative data analysis and validation.
Evidence: The instructions in SKILL.md and references/iteration-guide.md direct the agent to use the Bash tool to run voice-analyzer.py, voice-stylometry.py, and voice-validator.py.
Execution Details: Scripts are targeted at specific local paths such as ~/.claude/scripts/ or the project's scripts/ directory, minimizing the risk of arbitrary path execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process writing samples provided by the user, which creates a surface for indirect prompt injection attacks where the LLM might follow instructions embedded within the samples.
Ingestion points: Writing samples are read from the skills/voice-{name}/references/samples/ directory during the collection and analysis phases.
Boundary markers: While the skill suggests using horizontal rules (---) or headers to separate samples, it lacks robust sanitization to prevent the agent from interpreting sample content as instructions.
Capability inventory: The skill environment includes high-privilege tools like Bash, Write, and Edit, which could be targeted by a successful injection.
Sanitization: The skill mitigates this risk by including a mandatory Step 6: VALIDATE which uses voice-validator.py check-banned to scan for suspicious AI-generated patterns and tells.

create-voice