What is prompt engineering?
Prompt engineering is the practice of crafting effective instructions for AI models. It's both an art and a skill, and the good news is that anyone can learn it. You don't need a technical background or coding experience. You just need to understand how to communicate clearly with AI.
Think of prompt engineering like learning to work with a new colleague. At first, you might not know the best way to explain tasks to them. You learn through experience what level of detail they need, what kinds of instructions work best, and how to get the results you're looking for. Prompt engineering is the same process, except your colleague is an AI model.
💡Note
In the context of Tines, prompt engineering helps you get the most out of our AI features. Whether you're using AI to transform data, analyze information, or generate content, the quality of your prompts directly affects the quality of your workflows that use AI.
The tips included in this section are just general recommendations for prompt engineering. We’ll dig into more specific examples as we explore the different AI features in Tines.
Write clear and effective prompts
Clarity is the foundation of good prompt engineering. Here are the core principles that make prompts effective:
Be specific about your goal: Vague prompts lead to vague results. Instead of "Look at these security alerts," say "Identify which of these security alerts require immediate action based on their severity level."
Use action verbs: Start your prompts with clear action verbs such as "extract," "summarize," "list," "identify," "compare," or "categorize."
Define your output format: Tell the AI how you want the results structured. For example: "Return a bulleted list of the top five issues, with each item including the issue ID and a brief description."
Avoid ambiguous language: Replace subjective terms with specific values. Instead of "recent alerts," say "alerts from the past 24 hours." Instead of "a few examples," say "three examples."
Include constraints: State limits or boundaries upfront. For example: "Summarize this incident report in no more than three sentences. Focus only on the root cause and resolution."
🪄Tip
Provide details and examples
Once you have a clear prompt structure, adding context makes it even more effective. Here's how to provide the right details:
Explain your data and terminology: Define organizational terms and categories. For example: "In our system, P0 means complete service outage, P1 means major functionality impaired, P2 means minor issues, and P3 means feature requests."
Show the format you want: Include a sample of the desired output structure, like "Format each result like this: [Alert ID] - [Severity] - [Brief description]" or provide a JSON example with the exact fields you need.
Mention edge cases: Call out special situations the AI should handle. For example: "If the status field is empty, categorize it as 'unknown'. If the date is in the future, flag it as an error."
Provide context about your goal: Explain why you're asking for something when it helps. For example: "I need to send a summary email to management, so focus on high-level outcomes rather than technical details."
🪄Tip
Iterate and refine prompts
Even experienced prompt engineers rarely get it perfect on the first try. Iteration is a normal and valuable part of the process.
Start simple, then build: Begin with a basic version that addresses the core task. Test it, see what happens, then add detail based on what you learn. You might start with "Summarize this alert" and refine to "Summarize this alert in two sentences: one describing the issue and one describing the recommended action."
Test with real data: Use actual data from your workflows. Real data reveals edge cases and unexpected situations that test data might not show.
Adjust one thing at a time: When refining a prompt, change one element at a time when possible. This helps you understand what's actually making a difference.
Try different phrasings: Sometimes a small change in wording makes a big difference. "List all critical alerts" might give you different results than "Identify all critical alerts." Experiment with different ways of saying the same thing.
Know when to simplify: If your prompt is getting very long or convoluted, step back. Maybe you're trying to do too much in one prompt. Consider breaking the task into smaller steps.