The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes a mechanism for the agent to ingest interaction data and 'learn' new rules, creating a surface for indirect prompt injection. * Ingestion points: Interaction logs and tool execution outcomes (SKILL.md). * Boundary markers: The instructions mention 'misaligned-replay guards' and 'explicit validation' for security-sensitive flows, but do not define specific technical delimiters (SKILL.md). * Capability inventory: The agent is directed to perform file-writes to system_rules.md, AUDIT_LOG.md, and update an internal Knowledge Graph (SKILL.md). * Sanitization: The skill recommends a 'Data Loss Prevention (DLP) sanitization pass' to ensure ephemeral secrets are not permanently stored (SKILL.md).
[DYNAMIC_EXECUTION]: The skill facilitates the generation and adoption of new behavioral rules at runtime. By writing to system_rules.md, the agent modifies its own reasoning logic based on potentially untrusted runtime data.
[PERSISTENCE]: Modifications made to the agent's system_rules.md and 'Best Practices' bank are intended to be long-term, allowing learned behaviors to persist across future sessions.

learning-adaptation