The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection as it ingests untrusted data from external web searches (via WebSearch/Browser tools) and user-uploaded files located in /mnt/agents/upload/.
Ingestion points: Research data is gathered by sub-agents and stored in /mnt/agents/output/research/ as dimension reports (dim*.md), which are then read by writer agents.
Boundary markers: The orchestrator uses structured task prompts and outline excerpts to guide sub-agents, though explicit sanitization for instructions hidden within the research content is not detailed.
Capability inventory: Sub-agents have the ability to write files to the output directory and execute Python code via the IPython tool for chart generation.
Sanitization: The skill emphasizes the resolution principle where the orchestrator resolves all instructions before passing them to sub-agents, reducing the chance of autonomous sub-agent misinterpretation, but does not specify content filtering for ingested data.
[DYNAMIC_EXECUTION]: The skill utilizes the IPython tool to dynamically generate and execute Python code (specifically matplotlib and seaborn) for creating data visualizations during the content creation stage. It also dynamically constructs sub-agent system prompts by inlining content from domain-specific style files at runtime.
[EXTERNAL_DOWNLOADS]: The workflow depends on an external capability for document conversion located at the absolute path /app/.agents/skills/docx/SKILL.md. This is a dependency on a local system resource rather than a remote download.

report-writing