The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill dispatches agents to process untrusted text without explicit isolation, creating a vulnerability where malicious instructions in the text could influence agent behavior.
Ingestion points: External data is interpolated into the prose and craft reviewer agents via {generated_text} and {register_content} in SKILL.md.
Boundary markers: The skill lacks delimiters (e.g., XML tags) or instructions for the agents to ignore embedded commands within the text.
Capability inventory: In SKILL.md, the skill executes a subprocess via python plugins/imbue/scripts/citation_verifier.py and performs multiple file-write operations including saving the final text, writing findings to .review/findings.json, and saving snapshots to ~/.claude/voice-profiles/.
Sanitization: No input sanitization or validation of the external text is specified.
[COMMAND_EXECUTION]: The skill executes a local Python script as part of its verification workflow.
Evidence: The SKILL.md file includes a command to run python plugins/imbue/scripts/citation_verifier.py to validate citations in a findings file.

voice-review