The Agent Skills Directory

[PROMPT_INJECTION]: Indirect Prompt Injection Surface. The core functionality of the skill involves dynamically loading and applying external instruction sets ('skills', 'activities', 'processes') as behavioral overlays on characters and environments.
Ingestion points: The skill ingests external data through the MOUNT command, which references file paths (e.g., .moollm/skills/{skill-name}/skill-parameters.yml and activity files like debates/lincoln-vs-douglas.yml) to define new behaviors.
Boundary markers: There are no defined delimiters or 'ignore' instructions specified in the documentation to prevent the overlay instructions from overriding the agent's primary safety or identity constraints.
Capability inventory: The skill is granted read_file and write_file permissions to manage these behavioral states and load configuration data.
Sanitization: No explicit validation or sanitization process is described for the content of the mounted skills, meaning a malicious skill definition could contain instructions that redirect agent behavior or bypass expected constraints.
Identity Suppression: The documentation explicitly describes 'CATASTROPHIC' combinations that result in 'Identity Destruction' (e.g., suppressing the core traits of characters like Pee-wee Herman or Bob Ross). This encourages the agent to disregard its base instructions in favor of the mounted 'overlay'.

mount