The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes user-supplied UI descriptions to generate code, which represents a surface for indirect prompt injection.
Ingestion points: User-provided <description> inputs (SKILL.md, Phase 1).
Boundary markers: The agent is instructed to restate the brief before design begins, providing a basic confirmation step (SKILL.md, Phase 1).
Capability inventory: The agent can write new files in an isolated directory and integrate selected variants into production code (SKILL.md, Phase 4 and 6).
Sanitization: No explicit sanitization of the input description is performed within the instruction set.
[REMOTE_CODE_EXECUTION]: The skill employs dynamic code execution to render generated UI variants in a browser-based harness.
Evidence: The picker harness uses innerHTML to mount variant code into a stage element (PICKER.md).
Context: This execution is restricted to an isolated prototype surface to prevent impact on production systems, as mandated by the skill's hard rules (SKILL.md, Phase 4).

prototype