The Agent Skills Directory

[PROMPT_INJECTION]: The skill recommends a specific tone pattern in SKILL.md that explicitly tells the AI agent that user instructions should supersede its initial directives.
Evidence: "Instructions further down the conversation, including user's own, ALWAYS override prior style, tone, formatting, and initiative preferences."
This pattern is dangerous as it intentionally lowers the agent's resistance to malicious user prompts that seek to bypass original constraints.
[INDIRECT_PROMPT_INJECTION]: The skill provides architectures for agents to process untrusted data, specifically in the case studies for small models.
Ingestion points: small-models.md contains a case study for generating session titles based on data provided in a <user> tag.
Boundary markers: The skill recommends using explicit XML-style tags (e.g., <user>, <title>) to separate instructions from data, though these can still be bypassed in adversarial scenarios.
Capability inventory: This skill is documentation-based and does not define active tools, subprocesses, or network operations.
Sanitization: The skill correctly recommends post-processing (e.g., normalizeGeneratedTitle) in external code to strip suspicious content from model outputs, which is a valid mitigation strategy.

system-prompts