The Agent Skills Directory

[SAFE]: The skill implements proactive security measures based on OWASP Top 10 for LLM Applications and NIST AI RMF, focusing on preventing prompt injection and hidden context exposure.
[SAFE]: It mandates strict change-control workflows, including regression testing and red-teaming for all prompt updates before deployment.
[SAFE]: Secret management practices are enforced by explicitly forbidding API keys, credentials, or PII within prompt files, instead relying on pre-processors and authorized runtime controls.
[SAFE]: System message patterns include mandatory anti-jailbreak guards that instruct the model to treat external data as non-executable and to never reveal its internal instructions.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content such as sample prompts and retrieved knowledge chunks.
Ingestion points: Reads AI_Architecture_Spec.md, AI_Feature_PRD_Spec.md, and sample prompts as inputs to generate the specification (logic.prompt).
Boundary markers: Mandates the use of explicit delimiters like <document> and <tool_result> tags to isolate data from instructions (references/system-message-patterns.md).
Capability inventory: Limited to read/search access of authorized project artifacts and document validation checks; network access is restricted to verification of facts (SKILL.md).
Sanitization: Instructs the inclusion of "do-not-execute-instructions" clauses for all retrieved data blocks and mandates PII redaction by a pre-processor (SKILL.md).

13-ai-prompt-and-system-message-spec