The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user content, including research files and reference poems, and passes them to subagents for processing without explicit sanitization or boundary markers. \n
Ingestion points: User-uploaded files (copied to /mnt/agents/output/research/) and query-provided reference poems (referenced in task prompts). \n
Boundary markers: The instructions lack explicit delimiters or 'ignore embedded instructions' warnings when passing user data to subagents. \n
Capability inventory: The skill can spawn new agents (create_subagent), write files to /mnt/agents/output/, and execute local Python scripts. \n
Sanitization: No evidence of escaping or filtering of user-provided content before it is used in subagent instructions. \n- [COMMAND_EXECUTION]: The orchestrator invokes a local Python script via a subprocess to validate the output against requirements. \n
Evidence: Execution of python skills/general-writing/scripts/check_wordcount.py after poem generation. \n- [DYNAMIC_EXECUTION]: The orchestrator dynamically constructs system prompts and task instructions for its subagents by inlining external style guides and generated blueprints. \n
Evidence: Subagent system prompts for poetry_writer, form_checker, and sound_editor are built at runtime by inlining contents from ../anti-ai.md and the blueprint file agent.outline.md.

poetry