The Agent Skills Directory

[COMMAND_EXECUTION]: The skill's activation and completion sequences instruct the agent to execute a local Python script using uv run. Specifically, it calls {project-root}/_bmad/scripts/resolve_customization.py to handle configuration merging. While these scripts are located within the project's own directory structure, this represents a reliance on command-line execution for core functionality.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to read and process external data, such as Game Design Documents (GDD) and user-provided story concepts, which are then interpolated into documentation templates. This ingestion process lacks explicit sanitization or strict boundary delimiters, creating a surface where malicious instructions embedded in a GDD could influence the agent's behavior. The risk is partially mitigated by the skill's highly structured step-file architecture and clear facilitator role definition.
Ingestion points: The skill loads context from project files in SKILL.md (Step 3: Load Persistent Facts) and specifically searches for and extracts content from GDD files (*gdd*.md) in steps/step-01-init.md.
Boundary markers: The skill defines a strong persona as a "veteran narrative designer facilitator" and uses a "step-file architecture" to constrain the agent's flow, but it does not use explicit "ignore embedded instructions" delimiters when reading external content.
Capability inventory: The agent is authorized to execute local scripts via uv run, write to the filesystem (e.g., narrative-design.md), and read various project files (YAML, TOML, Markdown).
Sanitization: There is no evidence of specific input validation or sanitization for the narrative data before it is written to the output document or used in the workflow.

gds-create-narrative