The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill harvests lessons from current-session data or user-supplied evidence to update or create new skills, documentation, and Architecture Decision Records. This creates a potential surface where malicious session content could be persisted into the agent's instructions. * Ingestion points: Current-session evidence and user-supplied data in the 'Harvest' phase. * Boundary markers: Explicitly requires proposing a unified diff and waiting for user approval before writing to docs or ADRs. * Capability inventory: Ability to modify repository skill files, documentation, and ADR files. * Sanitization: Relies on a classification logic and a manual human-in-the-loop approval step to validate lessons before they are applied.

reflect