The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a robust mitigation strategy for indirect prompt injection. The script scripts/_untrusted_doc.py provides a wrap_untrusted_doc function that encapsulates project-authored content (like design briefs and wireframes) within unique delimiters (<<<UNTRUSTED_PROJECT_DOC>>>) and a warning to the agent to treat the content as data only. This is supported by explicit instructions in SKILL.md to disregard instructions or commands embedded in untrusted design inputs.
[DATA_EXFILTRATION]: The skill performs local analysis of Flutter code and design files. It does not contain any network operations targeting external domains. Network references are limited to localhost (127.0.0.1) for local visual auditing purposes, which is a safe and common development practice.
[COMMAND_EXECUTION]: While the skill involves running Python scripts for verification and handoff management, these scripts only use standard libraries (e.g., argparse, pathlib, re) and perform safe operations such as regex-based code auditing and file path resolution. There is no evidence of eval(), exec(), or unsafe subprocess calls on user-supplied data.

design-handoff-flutter