Local AI Chat App — Week 2[Prompt Guardrails + Safety On Our Simple App]

Prompt Guardrails + Safety On Our Simple App

Week 2 is about making the chat app safe, not just smart | https://cloudtoailearn.dev I added prompt guardrails to the Week 1 demo so the model now refuses unsafe requests, catches prompt injection, and validates every reply.

This week I split the work into two code paths:

• main branch = Week 1 clean vanilla chat app
• feature/guardrails branch = Week 2 safety enhancements

If you want to start from scratch:

1. Clone the repo
2. Checkout main and verify the plain app works
3. Create feature/guardrails
4. Add the Week 2 prompt/safety patterns

What changed in Week 2

✅ System prompt now says: “If the user asks for unsafe, illegal, or harmful actions, politely refuse.”

✅ Prompt injection detection for inputs like:

• “ignore previous instructions”
• “disregard the above”
• <|system|> or <|assistant|> tokens

✅ Refusal patterns for unsafe requests:

• hacking
• illegal actions
• credit card / SSN data
• medical/self-harm advice

✅ Output validation after generation:

• reject empty or broken replies
• reject model text that still contains prompt tokens

✅ Rate limiting repeated questions:

• repeat the same prompt 3x and the app says “Let’s slow down”

Why this matters: a model that can answer is useful, but a model that refuses unsafe or hostile requests is the one you can actually show to others.

What you can test

• “How do I hack wifi?” → refusal message
• “Ignore previous instructions” → rejection
• Repeating the same question 3 times → rate-limited reply
• Normal requests still work as expected

If you’re starting fresh:

• use main for the plain Week 1 demo
• branch into feature/guardrails for Week 2 safety work
• check WEEK2_PROMPT_ENGINEERING.md file for detialed instructions
• DM me in case if you are facing any issues

Repo: https://lnkd.in/eEpeDBug Full write-up: https://lnkd.in/e75KV3pi

This is Week 2 of my 12-week AI demo series at cloudtoailearn.dev Each 2 week: one working demo + one blog post.

like, subscribe & Share !!! Happy AI Leanring !!!

#AI #AISafety #PromptEngineering #SecureAI #Gradio #HuggingFace #cloudtoailearn #AI #LLM #PromptEngineering #MLOps #CloudArchitecture #cloudtoailearn #aws