The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data by reading target repository instructions, UI source code, and design tokens (SKILL.md). This attack surface is mitigated by the mandatory 'Plan Mode' constraint, which prevents automated execution, and the requirement for unambiguous human approval of the final text specification.
Ingestion points: Repository instructions, source code, design system documentation, and user-supplied screenshots.
Boundary markers: Instructions mandate grounding the proposal in authoritative contracts and strictly following iterative feedback rounds.
Capability inventory: The skill utilizes the $imagegen tool for local previews and delegates persistence to the $add-issue skill, which can only be loaded after a separate human request.
Sanitization: The workflow relies on manual human verification of the authoritative text specification before handoff to implementation tools.

redesign-ui