Instructions let you control AI agent behavior in ways beyond the default persona options. This article discusses best practices for creating effective instructions and gives you examples of good and bad instructions.
This article contains the following topics:
Related articles:
Best practices for creating effective instructions
To get the best results from your instructions, keep the following best practices in mind:
Write instructions in English
Instructions should always be created in English, regardless of any other AI agent language settings. Instructions written in English produce better results because the underlying AI model is currently better tuned for the English language.
Remember the purpose of instructions
Instructions allow you to refine an AI agent's responses beyond just tone of voice. They can enforce specific language, avoid certain phrases, or ensure key information is included in replies.
What instructions can’t do, however, is fundamentally change the content returned by the AI agent. For example, instructions cannot be used to search a different knowledge source or to cause an escalation.
Instructions are applied after the AI agent has understood the end user’s message, queried a knowledge source accordingly, and generated an answer ready to be sent back to the end user. At this point in the process, instructions can be used to shape—but not fundamentally change—the way that answer is presented to the end user.
Be clear and direct
When crafting instructions, clarity is key. Use straightforward language and direct commands.
For example, instead of saying:
- "I want the AI to always use our brand name."
Say this instead:
- "ALWAYS use 'Acme Labs™' instead of 'Acme.'"
A direct approach helps the AI understand and implement your instructions more effectively.
That said, expect your instructions and tone of voice settings to be interpreted as strong guidance rather than rigid commands. Due to the way generative AI works, your AI agents might not execute every instruction perfectly every time, so it’s important to maintain realistic expectations about their performance and test them thoroughly.
Use capitalization for emphasis
When you want to stress certain behaviors in your instructions, using capitalization can be effective. For example, instructing the AI agent to "ALWAYS include a link" will highlight the importance of that directive.
Create one instruction per directive
Instead of including multiple directives in a single instruction, break them down into separate instructions. Ensure that each instruction contains no more than a single directive. This approach:
- Enhances the likelihood that the AI agent will respect your instructions.
- Makes it easier for you to edit and manage instructions over time.
Avoid conflicting instructions
Ensure that your instructions don’t contradict each other. For example, don’t write an instruction that says to end all messages with a smiley face if you have another instruction that says not to use emojis.
Also, make sure your instructions don’t conflict with your chosen tone of voice. For example, if you configured the AI agent’s tone of voice to be informal, avoid instructions that suggest overly formal or abrupt responses. Consistency in messaging is crucial for a seamless customer experience.
Use existing features appropriately
Remember that instructions are not a substitute for other features. For specific tasks like customizing the standard responses or persona, use the designated features instead of relying on instructions.
In particular, take full advantage of the available persona settings, including business profile, tone of voice, and reply length. These features help define how your AI agent interacts with customers, enhancing the overall user experience.
Always test your instructions
Before making any instruction active, test the individual instruction and all your instructions together to see how the AI agent responds. Testing helps you identify any issues with wording or clarity. If an instruction doesn’t yield the expected outcome, try rephrasing it. The wording of an instruction plays a significant role in determining how an AI agent responds.
After publishing your AI agent, continue to monitor its interactions with end users. If certain instructions aren’t performing as expected, revisit and tweak them. The AI agent’s effectiveness can improve over time with careful adjustments based on real-world usage.
Examples of good and bad instructions
Below are examples of instructions, some good and some bad. Each example also includes an analysis of how following or ignoring the best practices above impact the instruction’s effectiveness.
Good instructions
-
Example: If the user asks any questions about how we are regulated,
then send the following response verbatim: “Thank you for reaching out to
us. Please see our regulatory information at
https://help-center-link.com/regulation.”
- Why it’s good: Clear and direct.
-
Example: ALWAYS say “Enjoy your flight!” at the end of your message
- Why it’s good: Uses capitalization for emphasis.
-
Example: When your answer is a list of step-by-step instructions,
always list them as 1, 2, 3 regardless of the formatting of the source.
- Why it’s good: Clear and direct.
-
Example: At FinanceCo, we cannot offer any financial advice ever.
Make sure you NEVER give the user financial advice of any kind. If you think
an answer could be construed as financial advice, make sure to add a
disclaimer at the beginning that says “Please note, we cannot provide
financial advice.” followed by a new paragraph.
- Why it’s good: Uses capitalization for emphasis. Though note that, as mentioned above, instructions cannot be 100% foolproof, so expect some potential instances of the AI agent not including the disclaimer when you might want it to.
Bad instructions
-
Example: You can use emojis, but only if they are more likely to be
appropriate than not. Decide yourself.
- Why it’s bad: The language is too vague and leaves too much up to interpretation.
-
Example: Always welcome the user.
- Why it’s bad: This conflicts with the greeting message feature.
-
Example: Always escalate to an agent if the user mentions being
hacked.
- Why it’s bad: Instructions cannot trigger escalations.