The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to generate 'Personality Instruction Blocks' which are interpolated with user-provided tasks for execution by other agents.
Ingestion points: Untrusted user tasks and personality requirements are processed in the SKILL.md workflow and workflow.md to create instructions.
Boundary markers: The skill explicitly instructs users to separate the 'Personality Instruction Block' from the 'Common Task' and 'Material/Permissions' blocks to prevent schema confusion.
Capability inventory: The generated instructions are intended to influence the target agent's internal reasoning process, attention, and decision-making order.
Sanitization: The skill contains robust 'Invariants' in SKILL.md that strictly forbid personality materials from rewriting task goals, facts, permissions, or safety standards. It also prohibits generated instructions from referencing the design skill itself to prevent circular logic or prompt leaks.
[SAFE]: All external URLs found in the skill (e.g., aclanthology.org, arxiv.org, doi.org, pubmed.ncbi.nlm.nih.gov) point to well-known academic repositories and research organizations. These are used solely for providing theoretical context and do not involve executable code or untrusted data ingestion.
[SAFE]: The 'Actor-style' and 'Induction' methods described in references/actor-induction.md and references/indirect-induction.md are documented as prompting techniques for controlled behavior modification. The skill includes specific warnings against using these methods to bypass safety filters or grant the agent unauthorized authority.

agent-personality