architect
Pass
Audited by Gen Agent Trust Hub on Sep 7, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill defines a structured architectural workflow including grounding, sketching, and evaluation phases. It uses external model runners to generate design candidates but does not perform any suspicious network or file system operations. It accesses local configuration files (~/.claude/pstack-models.md) to determine model preferences, which is consistent with its stated purpose.
- [INDIRECT_PROMPT_INJECTION]: The architectural workflow processes user-supplied tasks which are subsequently passed to multiple model runners during the sketch phase. 1. Ingestion points: User design requests and grounded system artifacts are ingested in SKILL.md (Phase B). 2. Boundary markers: Employs specific templates (references/rationale-template.md) and a dedicated runner prompt (references/runner-prompt.md) to structure interactions and constrain outputs. 3. Capability inventory: The skill is designed to produce code signatures, type definitions, and module structures in the local environment (Phase D and Outputs). 4. Sanitization: No specific content filtering is applied to the incoming architectural requirements beyond the provided templates.
Audit Metadata