General guidelines for writing instructions for generative AI large language models (LLMs)

  • Release version: Yokohama
  • Updated June 16, 2025
  • 7 minutes to read
    • Summarize
    Summarized using AI
    This content was generated using new OpenAI-powered functionality. Results are provided on an as is basis and are not guaranteed to be accurate or complete.

    Summary of General guidelines for writing instructions for generative AI large language models (LLMs)

    When using ServiceNow Now Assist products and skills, customers can provide specific instructions or guidance to generative AI large language models (LLMs). Writing effective instructions for LLMs differs significantly from traditional keyword searches because LLMs perform tasks based on the instructions rather than simply retrieving information. Clear, detailed, and actionable instructions help the LLM deliver precise and useful results tailored to the user's needs.

    Show full answer Show less

    Key Features

    • Instruction Components: Effective LLM instructions typically include four parts:
      • Goal: Define the desired outcome or result.
      • Context: Explain why the task is needed and who is involved.
      • Expectations: Specify how the LLM should fulfill the request, including tone and format.
      • Source: Indicate which information sources the LLM should use for answering.
    • Instruction Style: Use direct, imperative verbs and simple sentences without jargon or pronouns to avoid ambiguity and misinterpretation.
    • Task Tailoring: Customize instructions based on the type of task, such as simple search, information gathering, chat interaction, content creation, or workflow execution.
    • Contextual Clarity: Provide clear definitions and specify the target audience or user role (e.g., Admins, Builders, Agents, Requesters) to improve instruction accuracy.
    • Chain-of-Thought Instructions: Detailed, step-by-step reasoning within instructions helps LLMs produce more accurate and reliable answers.
    • Expectation Management: Customers can guide the style, tone, and formatting of responses, such as requesting professional but friendly language or bulleted lists for readability.
    • Source Configuration: Admins must configure access to relevant data sources like Microsoft Teams, SharePoint, incidents, or knowledge articles to enable LLMs and AI agents to retrieve accurate information.

    Practical Application for ServiceNow Customers

    ServiceNow customers using Now Assist can enhance the effectiveness of their generative AI interactions by:

    • Clearly defining the goal of their request with actionable verbs.
    • Providing sufficient background and user context to avoid ambiguities.
    • Specifying how they want the answer to be presented, including tone and format.
    • Testing and iteratively refining instructions to improve results as the LLM evolves.
    • Ensuring necessary data sources are configured by admins to enable comprehensive and accurate LLM responses.

    Additional Resources

    ServiceNow offers tailored guidelines for creating LLM instructions across various Now Assist skills and tools, including app generation, analytics, catalog item creation, code generation, flow generation, Virtual Agent topics, RPA bots, and UI generation. These resources help customers optimize instruction writing for specific use cases and maximize the value of generative AI within the ServiceNow platform.

    When using Now Assist products and skills, you may have the option to give specific instructions or other guidance to the LLM. Writing generative AI instructions is different from conducting a keyword search. Use the following general guidelines when crafting your instructions.

    Writing instructions for generative AI is very different from using search keywords. Keywords are the words that you might expect to appear in your results. For example, if you search for "gray bobtail cats," then you can reasonably expect your search results to return with topics or media that is about gray cats, bobtail cats, or even just cats in general. But with generative AI, you are asking the LLM to perform a task for you. The phrase "gray bobtail cats" does not include a verb to tell the LLM what to do. What about these gray cats? Should it locate all there is to know about them? Should it find gray bobtail cats to adopt? Should it create a picture of a gray bobtail cat? Should it be a realistic picture, or more of a line drawing? Generative AI needs more than just keywords.

    General LLM instructions

    Use instructions or questions to tell the LLM what you want. They can include four parts:
    Goal
    What kind of result do you want from the LLM?
    Example: I want a list of 3-5 bullet points to prepare me...
    Context
    Why do you need it, and who is involved?
    Example: ...for an upcoming meeting with [client], focusing on their current state and what they're looking to achieve with their "Phase 3+" brand campaign.
    Expectations
    How should the LLM best fulfill your request?
    Example: Please use simple language so I can get up to speed quickly.
    Source
    What information or other resources would you like the LLM to use?
    Example: Focus on email and Teams chats with [people] since June.
    Figure 1. Sample LLM instructions
    Instructions for the LLM that includes a goal, some context, your expectations, and sources to be used.

    Continually test and refine your instructions. Creating good LLM instructions is an iterative process, and as the LLM model learns, you may want to modify your instructions over time.

    Stating your goal

    When constructing an LLM description or instruction, consider these basic guidelines.
    • Lead with action verbs. Use the imperative form or direct commands.
    • Be direct and use simple sentences rather than complex ones.
    • Be specific.
    • Don't use jargon or slang terms.
    • Avoid references to third parties or pronouns. Removing the subject or any identifiers generally prevents the LLM from personifying or otherwise misidentifying the end user.
    • Your words instruct the logic that generative AI will use. Detailed, chain-of-thought instructions work well for this.

    Instructions should also be tailored to the type of task. The following table describes the different kinds of tasks and the sort of instructions you might write for each circumstance.

    Table 1. Types of generative AI tasks and example instructions
    Task type Description Example instruction
    Simple search Simple search for an answer. When is the next company holiday?
    Answer Gather information from multiple sources and provide a summarized answer. What were the major customer support issues in the past 30 days?
    Chat A back-and-forth conversation in which the LLM is getting additional information from the requester. I have a new phone and now I can’t access Okta.
    Create Create a new ServiceNow component. Write a new KB on common reasons for slow query execution and how to fix it, based on problems created in the last 12 months.
    Workflow Leverage existing workflows and create conversations from them. Reset my Okta password.
    Figure 2. LLM instruction workflow
    Different kinds of LLM task instructions can be provided to the LLM, which then displays results to the requester. The types of instruction include answers, workflows, search, chat, and create.

    Providing context in your instructions

    Providing context to the LLM may feel like you are stating the obvious. For example, you may need to explain why your user would want to perform the task, or explain more about what the task is about. If you're using language that could have alternative meanings, you may want to define your terms. For example, if your instructions are about Microsoft Teams, you may need to say something like, "Teams refers to Microsoft Teams, an application that employees in a company can use to communicate with one another individually or in groups."

    When providing context, think about the target audience for the task. This will help you to write better instructions. Mentioning whom the task is for also helps the LLM carry out the task.

    Types of users to consider:
    Admins
    Admins configure skills in the Now Assist Admin console. They work with platform owners and product owners for tasks and requirements. Subject-matter experts check the accuracy of generative AI results. Governance boards or committees may oversee final sign-off on the skill.
    Builders
    Builders create assets such as applications and workflows. Their skill level may vary from no-code, low-code, mid-skill, or high-skill. They mostly interact with each other and admins.
    Agents
    Agents provide technical assistance, customer support, or other problem-solving help for users of a product, service, or organization. Agents work with their peers and support many kinds of requesters.
    Requesters
    Requesters may include your organization's employees, partners, or customers. Requesters encounter generative AI mostly in a self-service context. Generative AI provides the opportunity to requesters to solve the problem themselves before they resort to asking an agent for help.

    In addition, you may need to "instruct" the LLM how to arrive at the kind of answer you want. Detailed chain-of-thought instructions are more likely to produce the results you're looking for. If you make assumptions in your instructions, your results may vary wildly.

    Table 2. Example of chain-of-thought instructions
    Standard instructions Chain-of-thought instructions

    Q: Start with 5 tennis balls. Buy 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls are there now?

    A: The answer is 11.

    Q: The cafeteria had 23 apples. If the cafeteria used 20 to make lunch and bought 6 more, how many apples does the cafeteria have?

    Q: Start with 5 tennis balls. Buy 2 more cans of tennis balls. Each can has 3 tennis balls. How many tennis balls are there now?

    A: There are 5 balls. 2 cans of 3 tennis balls each is 6 tennis balls. 5+6=11. The answer is 11.

    Q: The cafeteria had 23 apples. If the cafeteria used 20 to make lunch and bought 6 more, how many apples does the cafeteria have?
    LLM results (incorrect!) LLM results (correct!)
    A: The answer is 27. A: The cafeteria had 23 apples originally. They used 20 to make lunch. So they had 23-20=3. They bought 6 more apples, so they have 3+6=9. The answer is 9.

    Expectations

    You can specify how you want your results to sound and look. For example, you can tell the LLM what kind of tone you want it to use.

    Table 3. Modifying the tone of LLM results
    Instruction LLM result Evaluation of result
    Greet the user with a casual, laid-back tone. Hey there! (: Welcome to May 22, 2024, a Wednesday. Chill and have a great day! Too casual!
    Greet the user with a professional, business-like tone. Good day, welcome to our business. How may I assist you today? A little too formal.
    Greet the user with a professional, but friendly tone. Hello there! I'm glad to assist you today, May 22, 2024. How can I help you? Great, but I don't want the date.
    Greet the user with a professional, but friendly tone. Do not include the date.

    Hello! I'm here to help. How can I assist you today?

    		 Perfect.
    Instruction LLM result Evaluation of result

    In some Now Assist applications, you may have the option to specify formatting. In those cases, you can tell the LLM to provide answers in a bulleted list, for example. Bulleted lists are often easier to read.

    Figure 3. Enabling bulleted list results in the chat summarization skill
    In the Now Assist Admin console, some skills allow you to specify formatting. The chat summarization skill has an option to render results in a bulleted list.

    Source

    You can suggest a variety of sources that the LLM should use to find answers, including Microsoft Teams conversations, Microsoft SharePoint Online sites, incidents and cases, and internal knowledge articles. In order for an agent or skill to access all of these sources, an admin must configure access on the instance.

    For developers, the sources that a skill or AI agent can access may vary depending on the desired outcome. For example, the incident summarization skill uses the Incident table as its source. AI agents use different tools and knowledge sources, customized for the task they perform.

    For more information about configuring sources for skill or AI agent use, see the following topic areas:

    Additional guidelines for Now Assist skills and tools

    Table 4. Resources for writing LLM instructions for Now Assist skills
    Skill Reference
    App generation General guidelines for using Now Assist for app generation in ServiceNow Studio
    Analytics generation Guidelines and example questions
    Catalog item generation How to describe a catalog item
    Code generation General guidelines for code generation
    Flow generation
    LLM topic skill for Virtual Agent LLM description and instruction guidelines for Virtual Agent topics
    Now Assist Skill Kit General guidelines for Now Assist Skill Kit
    RPA bot generation General guidelines for RPA bot generation
    Test generation Design considerations for prompting
    UI generation General guidelines UI generation