The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains literal strings associated with prompt injection and jailbreak techniques, such as 'Ignore previous instructions', 'DAN', and 'reveal your instructions'. These are clearly documented as test cases within an educational framework for building robust applications and are not intended to hijack the agent.
[INDIRECT_PROMPT_INJECTION]: The skill provides patterns for processing external, untrusted data which presents an attack surface for indirect prompt injection. * Ingestion points: Data enters through placeholders like {{review}} and {{document}} in SKILL.md and test suites in references/evaluation-frameworks.md. * Boundary markers: The skill explicitly recommends using XML delimiters like <user_message> to isolate untrusted content in references/system-prompts.md. * Capability inventory: Included code snippets demonstrate the use of LLM APIs for processing external data in references/evaluation-frameworks.md. * Sanitization: Comprehensive guidance on injection defense, instruction hierarchy, and output guardrails is provided in references/system-prompts.md.
[METADATA_POISONING]: The skill's YAML frontmatter identifies a third-party GitHub profile as the author, which differs from the provided author context. This is likely an artifact of the skill being compiled from open-source references and does not appear to be a malicious attempt to deceive.

prompt-engineer