Improve AI Agent results with success criteria and examples

Beyond the essentials

You’ve learned to write clear prompts with action, content scope, and intent. These essentials work well for straightforward tasks: finding content, moving items through workflows, assigning contributors. But some tasks require more guidance. Consider this prompt: “Translate the Getting Started Guide content item to a new German language variant.“ The AI Agent will translate it. But will it keep your technical terms in English? Will it match the friendly, instructional tone of your other German content? Will it maintain consistency with how you’ve localized product names? Without additional guidance, the agent makes reasonable assumptions. Sometimes those assumptions match your needs. Often, they don’t quite hit the mark. That’s where quality boosters come in.

When to use quality boosters

Add success criteria and examples when the output quality, style, or format matters. Here’s how to tell if you need them. You need quality boosters when:

• You’re generating content at scale
  • Generating content (SEO metadata, summaries, descriptions)
  • Translating content while keeping the tone and terminology consistent
• Making decisions that could vary, such as taxonomy term assignment or appropriate phrasing
• Quality really matters and you want the best first-attempt results
• Any task where the gap between “technically correct” and “actually useful” is meaningful

The essentials alone work fine when:

• Finding or listing content (discovery tasks)
• Moving content through workflows (procedural tasks)
• Making clear-cut changes with no room for interpretation

Think of it this way: if you’d need to give another person detailed instructions to do the task right, the AI Agent needs those same details.

Principle 4: Define success

Success criteria tell the AI Agent what “good” looks like for your specific situation. They transform “generate something” into “generate something that meets these specific requirements”.

Types of success criteria

In a nutshell, think about what would make you reject a result and turn that into a positive requirement.

Translation prompt: Success criteria

Let’s build a translation prompt step by step. 1. Start with just the essentials: “Translate the Getting Started Guide content item to a new German language variant.“ The AI Agent will translate it. You’ll get German text. But what will that German text look like?

• The tone might be formal when your brand is conversational
• Technical terms might be translated when you wanted them in English
• Product names might be localized inconsistently
• The style might not match your other German content

2. Add your first success criterion - technical terms: “Translate the Getting Started Guide content item to a new German language variant. Keep technical terms like API and webhook in English.“ Better. Now AI Agent knows not to translate those terms. But tone and style are still open to interpretation. 3. Add tone requirements: “Translate the Getting Started Guide content item to a new German language variant. Keep technical terms like API and webhook in English. Match the friendly, instructional tone of our existing German content.“ Now you’ve defined the technical requirements and the voice. But “friendly, instructional tone“ is still somewhat subjective. What does that actually mean in practice? This is where you need the second quality booster, providing the agent with an example.

Principle 5: Show an example