The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains instruction bypass patterns, such as "Ignore previous instructions and instead...", in the references/synthetic-data.md file. These are explicitly provided as examples of adversarial inputs for testing LLM robustness and are not instructions directed at the agent itself.
[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for generating synthetic test data that includes adversarial payloads and encoding tricks (zero-width characters, homoglyphs). This creates a vulnerability surface where the agent might process these generated inputs during validation steps. However, the skill provides mitigation strategies including structured data formats and safety guardrails.
Ingestion points: The references/synthetic-data.md file identifies placeholders and variables that accept dynamic, potentially adversarial input.
Boundary markers: The skill advocates for structured outputs (JSON/XML) and the use of explicit delimiters like <thinking> and <answer> tags to separate logic from content.
Capability inventory: The skill involves LLM evaluation and prompt validation workflows, but does not include scripts with shell execution, file system writes, or network access.
Sanitization: Includes specific verification steps in references/techniques.md to check for harmful instructions and personal information before generating responses.

prompt-engineering