The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted external content, including URLs, pasted text, and screenshots, which serves as a potential vector for indirect prompt injection if the source material contains malicious instructions. The skill proactively mitigates this risk through several design choices: 1. Ingestion points: Defined in SKILL.md (Section 1), allowing for arbitrary text and web content. 2. Boundary markers: The skill requires explicit citations, tagging every claim as either [documented] (cited to source) or [inferred] (based on reasoning), as seen in SKILL.md (Section 3). 3. Capability inventory: The skill generates structured text reports and can write files to the filesystem, but file writing is strictly gated by explicit user consent and warnings regarding sensitive locations as per intake/VALUES-INTAKE.md and reference/DATA-HYGIENE.md. 4. Sanitization: The instructions include strong prohibitions against bluffing about unreadable sources, drafting replies for the user, or providing conclusions, ensuring the AI maintains a passive role as a 'mirror'.
[SAFE]: The skill implements a robust 'Judgment Contract' and 'Data Hygiene' framework that prioritizes user privacy and safety. It explicitly prohibits clinical advice, commercial solicitations, and unauthorized data persistence. It mandates user consent before any file operations and provides clear warnings when saving to potentially unsafe locations such as git repositories or shared drives.

what-do-i-make-of-this