The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data sources such as experiment logs (CSV, JSON, TSV) and training curves to generate reports and figures. This represents a standard attack surface for indirect prompt injection if an attacker-controlled log file contained malicious instructions. However, this is an inherent risk of data analysis tasks.
Ingestion points: Metric tables (CSV, JSON, TSV), training logs, and checkpoint metadata identified in SKILL.md.
Boundary markers: The skill does not explicitly instruct the agent to use delimiters or ignore embedded instructions within the processed data.
Capability inventory: File reading (logs), file writing (report generation), and figure generation (plotting via environment tools).
Sanitization: No explicit sanitization or validation of the content of the logs (beyond statistical validity checks) is mentioned.
[COMMAND_EXECUTION]: The skill instructs the agent to 'generate real scientific figures' using data and logs. This implicitly requires the use of plotting libraries (e.g., matplotlib, seaborn) within the agent's execution environment. This is the primary intended function of the skill and is documented as such in the USAGE.md and references/visualization-best-practices.md files.

results-analysis