The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill identifies design limits by reading and analyzing external data sources such as evaluation reports, error analyses, and run logs.
Ingestion points: The skill ingests data from external files including evaluation reports, error analyses, run logs, and research records (SKILL.md).
Boundary markers: There are no explicit instructions or delimiters defined to prevent the agent from following malicious instructions potentially embedded in these external logs or reports.
Capability inventory: The skill possesses the ability to write files (for recording exploration and creating prototypes) and can invoke multiple other specialized skills like ml-experiment and evaluation-design.
Sanitization: No mechanisms for sanitizing or filtering the content of the ingested reports are mentioned.
[DYNAMIC_EXECUTION]: The skill encourages the creation and execution of "throwaway prototypes"—short-lived code snippets intended to answer specific design questions.
Evidence: Instructions in references/throwaway-prototype.md guide the agent to write and execute code (e.g., state models, scoring rules) to observe behavior.
Mitigation: The skill mandates safety protocols for these prototypes, requiring they use synthetic data, avoid production endpoints, and run exclusively in a local sandbox.

design-exploration