The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a continuous learning pipeline that extracts patterns and solutions from session logs (session_log in SKILL.md).
Ingestion points: The agent context absorbs session logs and user-provided inputs as defined in instructions.md and the ConstructionLearningPipeline class.
Boundary markers: While the _validate_pattern method in SKILL.md checks for confidence scores and code validity, it does not include explicit delimiters or instructions to ignore embedded malicious prompts within the session data.
Capability inventory: The skill possesses filesystem and network permissions (claw.json), allowing it to persist learned patterns to a local knowledge base and interact with external project management APIs.
Sanitization: The skill performs basic structure validation and code snippet checking but lacks comprehensive sanitization for natural language instructions that could be re-injected into future sessions.
[COMMAND_EXECUTION]: The skill includes operational commands such as /learn, /add-pattern, and /export-kb (SKILL.md), which trigger the execution of file system and data processing logic.
[DATA_EXFILTRATION]: The skill includes functionality to export its internal knowledge base (/export-kb) and provides patterns for integrating with external APIs like Procore (SKILL.md). These operations handle project-specific data, though they align with the skill's stated purpose of knowledge preservation.

continuous-learning-construction