The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes local utility Python scripts via uv run across its lifecycle steps (e.g., {project-root}/_bmad/scripts/memlog.py, resolve_customization.py, resolve_config.py). It also executes platform-specific document opener utilities (open, xdg-open, start "") inside references/validate.md to display HTML reports. These are normal operational components of the BMad methodology structure.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted third-party inputs during the Discovery phase.
Ingestion points: Inputs are fetched during the 'Brain dump' stage (SKILL.md) where the user supplies product briefs, research materials, customer transcripts, and competitive analysis documents via copy-paste or local file paths.
Boundary markers: Missing or weak boundary delimiters are used during prompt interpolation when compiling these external texts into the context.
Capability inventory: Uses uv run to manipulate local file-logs via memlog.py and opens synthesized validation files in the browser using dynamic platform tools (open, xdg-open, start) within references/validate.md.
Sanitization: There are no specialized text cleaning or instruction filtering layers before the inputs are handed off to subagents for data extraction.

bmad-prd