atelier
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONEXTERNAL_DOWNLOADS
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill serves as a pipeline orchestrator that processes external, potentially untrusted data to drive the actions of downstream agents.
- Ingestion points: The skill ingests aesthetic direction from external files like
direction.md(via the Vision agent), user-provided briefs, and design data extracted from Figma files via theFramedelegate. - Boundary markers: It utilizes the
DESIGN_INTENT_HANDOFFstructured schema to define instruction boundaries and standardized metadata when delegating tasks to other agents. - Capability inventory: The skill orchestrates high-privilege agents including
Artisan(responsible for production frontend implementation) andForge(responsible for rapid prototyping). It also performs file system operations within the.agents/directory to maintain a persistent design system. - Sanitization: It employs structured handoff bundles and specific operation layers (such as parametric sliders and structured comments) to communicate intent, which provides more structure than free-form natural language prompts.
- [EXTERNAL_DOWNLOADS]: The skill references official resources from a well-known technology provider.
- Evidence: Includes links to Anthropic's official newsroom (e.g.,
https://www.anthropic.com/news/claude-design-anthropic-labs) to provide context for its design workflow philosophy.
Audit Metadata