The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process untrusted external data in the form of pasted memory exports. It mitigates this risk through strict boundary instructions that mandate the agent treat the paste solely as data and ignore any embedded instructions or directives. It includes a mandatory evidence chain:
- Ingestion points: User-provided text pastes (Flow Step 1).
- Boundary markers: Explicit instructions to ignore instruction-like content and treat input as data.
- Capability inventory: Restricted to memory tools only; network operations and external file writes are prohibited.
- Sanitization: A comprehensive privacy filter (Step 3) and summarizing prose into single-fact lines.
[PROMPT_INJECTION]: Deterministic detectors may flag phrases regarding ignoring instructions. In this skill, these are defensive measures instructing the agent to disregard potential injection attempts contained within the external data export rather than attempting to override the agent's core safety guidelines.
[DATA_EXFILTRATION]: The skill includes strong prohibitions against exfiltration, explicitly forbidding the fetching, following, or reproduction of URLs and links found within the imported data.

import-memory