Skip to main content
Practical Guidance

Best Practices

Practical guidelines for using AI effectively, safely, and responsibly.

Quick Checklist

Define clear success criteria before starting
Test with diverse inputs and edge cases
Verify facts from authoritative sources
Consider who might be affected by AI outputs
Document your AI usage and decisions
Stay updated on model limitations and updates
Gather and act on user feedback
Have human oversight for high-stakes decisions

Prompt Engineering

Learn more

Be Specific and Explicit

Instead of "summarize this", say "summarize this in 3 bullet points focusing on key findings".

DO

Write a 200-word explanation of transformers for a college freshman studying CS.

DON'T

Explain transformers.

Use System Prompts Wisely

Set context and constraints at the beginning to guide model behavior throughout.

DO

You are an expert Python developer. Always include error handling and type hints.

DON'T

Be helpful.

Chain-of-Thought Prompting

Ask the model to think step-by-step for complex reasoning tasks.

DO

Let's solve this step by step. First, identify the variables...

DON'T

What's the answer?

Provide Examples (Few-Shot)

Show the model what good outputs look like before asking for new ones.

DO

Example: Input: "happy" → Output: "joyful, elated, pleased". Now do: "sad"

DON'T

Give me synonyms.

Safe & Responsible Use

Learn more

Verify Critical Information

AI can hallucinate facts, especially dates, statistics, and technical details.

DO

Cross-reference AI outputs with primary sources for important claims.

DON'T

Blindly trust AI for medical, legal, or financial advice.

Understand Limitations

Know what AI is good at (pattern matching, synthesis) and bad at (novel reasoning, current events).

DO

Use AI as a starting point or brainstorming partner.

DON'T

Expect AI to replace domain experts or make final decisions.

Protect Privacy

Don't share sensitive personal data, trade secrets, or confidential information with AI systems.

DO

Anonymize data before using AI for analysis.

DON'T

Paste full customer records or proprietary code into public AI tools.

Consider Bias

AI models reflect biases in training data. Review outputs critically.

DO

Ask for multiple perspectives and test with diverse inputs.

DON'T

Use AI outputs for hiring or evaluation without human review.

Working with Code

Learn more

Review Generated Code

AI code often works but may have security flaws, inefficiencies, or subtle bugs.

DO

Test AI-generated code thoroughly; understand what it does before using.

DON'T

Copy-paste code directly into production without review.

Provide Full Context

Share relevant code, error messages, and expected behavior.

DO

Here's my function [code], the error [error], and I expect [behavior].

DON'T

"My code doesn't work, fix it."

Use for Learning, Not Replacing

AI is best as a teaching tool to understand concepts, not just get answers.

DO

Ask "explain why this approach works" or "what are alternatives?"

DON'T

Ask for code without trying to understand the solution.

Iterate and Refine

Treat AI as a collaborative partner—refine outputs through conversation.

DO

"Good start, but can you add error handling for edge case X?"

DON'T

Accept the first output as final.

Building with AI

Learn more

Start with RAG

Retrieval-Augmented Generation grounds AI in your data, reducing hallucinations.

DO

Index your knowledge base and retrieve relevant context before generation.

DON'T

Rely solely on fine-tuning for domain-specific knowledge.

Implement Guardrails

Add validation layers to catch inappropriate or incorrect outputs.

DO

Use content filters, output validators, and human review for critical paths.

DON'T

Deploy AI directly to users without safety checks.

Monitor and Evaluate

Track AI performance over time with real user interactions.

DO

Log inputs/outputs, measure quality metrics, gather user feedback.

DON'T

Deploy and forget—AI behavior can drift or degrade.

Design for Failure

AI will sometimes fail. Build graceful fallbacks.

DO

Have human escalation paths and honest "I don't know" responses.

DON'T

Promise 100% accuracy or hide AI limitations from users.

The Core Philosophy

AI is a powerful tool, not magic. Treat it like a brilliant but occasionally confused intern: it can do amazing work, but needs clear instructions, careful supervision, and fact-checking. The best results come from human-AI collaboration, not blind automation.