Day 5: Safety & Production

Today covers what happens when prompts leave your laptop. Prompt injection is the most underestimated attack surface in LLM applications. Guardrails catch unsafe outputs before users see them. Version control and CI/CD treat prompts as code. And at scale, prompt management becomes its own engineering discipline.

Lessons

1. Prompt Injection Defense , direct, indirect, and extraction attacks, and the defense stack
2. Jailbreak Prevention , role-play, encoding, and incremental escalation attacks
3. Content Safety & Guardrails , output filtering, classification, and safety layers
4. Prompt Versioning & CI/CD , treating prompts as code with version control, testing, and deployment
5. Prompt Engineering at Scale , prompt libraries, team workflows, and organizational standards

Next: Capstone: Build a Prompt Library