The Agent Skills Directory

[SAFE]: The skill is designed for structural knowledge management and data anonymization. It does not contain any executable scripts, network requests to untrusted domains, or attempts to bypass safety filters.
[DATA_EXPOSURE]: The skill includes robust privacy protocols in references/privacy-and-redaction-rules.md and protocols/material-boundary-protocol.md. These instructions mandate multiple levels of data redaction (L1-L3), ensuring that sensitive details like names, addresses, and private contact information are generalized or removed before the data is processed into cases.
[PROMPT_INJECTION]: The instructions in SKILL.md and the various protocols are purely functional, focusing on task execution and output formatting. There are no attempts to override the system prompt or ignore safety guidelines.
[INDIRECT_PROMPT_INJECTION]: The skill acts as a sink for external data (chat logs, etc.), which is a known attack surface. However, it mitigates this by enforcing strict 'fact boundaries' and 'interpretation' layers defined in protocols/material-boundary-protocol.md. It explicitly warns against treating model inferences or source claims as absolute facts, providing a defensive layer against embedded instructions in user-provided materials.

crossframe-casebook