The Agent Skills Directory

[DATA_EXPOSURE]: The researcher role uses tools like Glob and Grep to scan the project directory for existing interactive components and browser API usage. This is a legitimate functional requirement for local code analysis and does not involve sensitive file paths or network exfiltration.
[COMMAND_EXECUTION]: The coordinator role executes a single benign bash command pwd to resolve the current workspace root and establish session directory paths.
[INDIRECT_PROMPT_INJECTION]: The skill ingests user-provided task descriptions which are propagated to sub-agents (Designer, Builder). This attack surface is mitigated by strict coding constraints defined in the skill (e.g., zero-dependency rule, prohibition of eval/innerHTML, and requirement for vanilla JS only) which limit the impact of any potentially injected instructions in the generated code.

team-interactive-craft