The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill suite frequently handles untrusted external data (e.g., website content via agent-browser, wiki sources, and customer research data). The instructions include robust mitigations, explicitly directing the agent to treat fetched text, source-embedded commands, and model outputs as 'untrusted evidence' rather than instructions that could expand its task or authority.
[EXTERNAL_DOWNLOADS]: The skill references several external tools and packages, including agent-browser (from Vercel Labs, a well-known service), and various @andrue/* and @better-typescript/* packages which are owned by the skill's author. These are documented neutrally and align with the skill's technical purpose.
[COMMAND_EXECUTION]: Instructions guide the agent to use CLI tools for project management and software verification. These procedures include safety checks, such as requiring the agent to inspect commands and called scripts before execution and to use authorized, disposable environments for installations.
[DATA_EXFILTRATION]: The skill provides explicit guidance on secret management, instructing the agent to keep credentials, session secrets, and private data out of logs, artifacts, and version control. It emphasizes using secure secret requests rather than putting sensitive information in chat.
[SAFE]: No patterns of prompt injection, obfuscation, or persistence mechanisms were detected. The skill's structure and instructions are consistent with high-quality, security-conscious engineering practices.

help