The Agent Skills Directory

[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface in the 'Stakeholder Review' section. It ingests discovery synthesis data and persona-specific JSON files which are then interpolated into prompts for parallel task agents. While the prompts use clear instructional blocks, they lack strict boundary markers or sanitization for the external content.
Ingestion points: Ingests data from discover/synthesis.md and personas/ directory files.
Boundary markers: Uses thematic headers (e.g., 'FILES TO READ', 'INSTRUCTIONS') but lacks formal XML or high-entropy delimiters for interpolated content.
Capability inventory: The skill has access to Bash (for local state scripts), Write, and Edit tools for project file modification.
Sanitization: No explicit validation or escaping of the research content is performed before prompt construction.
[COMMAND_EXECUTION]: The skill executes a local shell script (update-phase.sh) located in the plugin's script directory to manage project state. This is a standard operational pattern for this environment and does not involve untrusted input in the command string beyond the project directory path.

consulting-define