The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to handle untrusted data, such as user inputs and retrieved context, which constitutes an indirect prompt injection surface. The skill proactively addresses this by providing extensive instructions on mitigation strategies.
Ingestion points: Data enters the context via $ARGUMENTS in SKILL.md, and through placeholders like <user_input> and <retrieved_context> described in the rule files.
Boundary markers: The skill explicitly recommends using XML tags and structural delimiters to separate instructions from data.
Capability inventory: The skill itself is purely advisory and contains no executable scripts or tool configurations.
Sanitization: The instructions recommend using specialized input classifiers (e.g., Llama Guard, ShieldGemma) and output validators to sanitize and verify content.
[PROMPT_INJECTION]: A deterministic detector flagged the phrase 'Ignore all previous instructions' in rules/memory-and-state.md. Upon manual review, this is a false positive. The phrase is used within a security guideline that instructs developers to 'reject content with embedded instructions' as a defense mechanism, rather than attempting to bypass safety filters itself.
[EXTERNAL_DOWNLOADS]: The skill contains numerous URLs to external documentation and research in references/primary-sources.md. All referenced domains (e.g., anthropic.com, openai.com, arxiv.org, hamel.dev) are well-known, reputable sources for AI engineering practitioners.

ai-engineering