The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains robust defensive instructions specifically designed to prevent the AI from following malicious commands that might be embedded in user-uploaded documents (e.g., CVs, academic papers, or USCIS notices). It explicitly instructs the agent to treat these files strictly as data and to ignore any text that attempts to override its instructions.
[INDIRECT_PROMPT_INJECTION]: The skill manages the risk of processing untrusted external data by defining clear boundary markers. It mandates that all ingested content be treated as data only and provides specific sanitization instructions to disregard any embedded prompts. This effectively mitigates the attack surface presented by the processing of third-party academic and legal documents.
[EXTERNAL_DOWNLOADS]: The skill automates the retrieval of official immigration forms from highly trusted government domains, including uscis.gov and dol.gov. It also fetches scholarly metadata from openalex.org, an established academic API. These operations are performed via transparent, standard Python scripts provided within the skill package.
[DATA_EXFILTRATION]: The skill implements a privacy-preserving 'local case folder' model. It ensures that sensitive personal identifiers (such as Social Security Numbers and Alien Registration Numbers) are stored only within a user-controlled local directory. The instructions strictly forbid referencing any files outside this self-contained folder, preventing accidental data exposure.

niw-petition