The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's workflow involves reading and processing external design documents (GDDs) and engine references, which introduces a surface for indirect prompt injection attacks.\n
Ingestion points: Design files in design/gdd/, engine references in docs/engine-reference/, and project documentation in .claude/docs/ (as seen in references/claude-skill.md).\n
Boundary markers: The skill does not explicitly define delimiters or instructions to ignore embedded prompts within the ingested markdown files.\n
Capability inventory: The skill utilizes Read, Write, Bash, and Task (agent spawning) tools across various phases of the architecture blueprint generation.\n
Sanitization: There is no explicit validation or sanitization step for content extracted from the GDDs before it is used to influence architectural decisions or document generation.\n
Mitigation: The workflow implements a strict human-in-the-loop protocol, using AskUserQuestion to obtain explicit user approval for each layer map, ownership definition, data flow, and final document write.

create-architecture