The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user-supplied text and analysis reports which are interpolated directly into the system prompt, creating a vulnerability to indirect prompt injection where instructions embedded in the input could influence agent behavior.
Ingestion points: The variables {text} (original writing) and {analysis} (pattern reports) are defined as inputs to the agent instructions.
Boundary markers: The Prompt Template lacks robust delimiters or XML tags to separate instructions from user-provided content, relying on simple labels like [Original Text]: which are easily subverted.
Capability inventory: The skill includes instructions to read local configuration files (.research/decision-log.yaml) and execute custom platform functions (diverga_check_prerequisites, diverga_mark_checkpoint).
Sanitization: No input sanitization, filtering, or specific instructions to ignore embedded commands are present in the provided templates.
[COMMAND_EXECUTION]: The skill specifies the use of custom platform-specific tools to manage its execution flow and state.
Evidence: Instructions such as diverga_check_prerequisites("g6") and diverga_mark_checkpoint("CP_HUMANIZATION_VERIFY", decision, rationale) trigger functional operations within the agent environment.

g6