General guidelines for code generation

  • Release version: Yokohama
  • Updated January 30, 2025
  • 1 minute to read

    • Use these general guidelines for code generation to get better code suggestions and create useful and accurate scripts.

    Writing prompts

    Write clear and specific but concise prompts
    Specify the expected outcome and context, including necessary details like task requirements, specific APIs if you know them, and any constraints.
    Experiment with different prompts

    As you refine and experiment, the Now LLM Service uses this feedback to learn and improve.

    • Try adjusting task instructions and incorporating examples, and then observe how code suggestions differ with different prompt styles and levels of detail.
    • Try including a short code snippet as an example of how to start the script with a single-shot prompt.
    • Track your prompts, including any modifications, and instructions for generating prompts to meet your specifications. This tracking enables easy regeneration of past results for comparative analysis.
    Character limit of the prompts

    Short and concise prompts generate better outcomes.

    On reaching 200 characters, a message appears to inform you that short, focused, task-oriented directions yield the best results.

    Input beyond 300 characters isn’t allowed.

    Table 1. Example prompts for code generation
    Strong prompt Weak prompt Notes
    Get incidents with related tasks Get incidents with tasks

    Includes sufficient detail.
    Use Glide aggregate to count number of P1 incidents closed between March 3 to April 13 assigned to admin Count P1 incidents between 3-3 and 4-13

    Includes the API name and more specific language.
    If open change request is P1, don’t allow reducing the severity unless it's the creator Don’t allow changing P1 change requests

    Includes more specific instructions on what shouldn't change.
    Glide record of the most recent change Latest change

    Includes the API name and more specific language.

    Reviewing code

    Review code
    Implement strict and detailed reviews of the AI-generated code to determine its accuracy, efficiency, and how well it adheres to your coding standards.
    Test code
    Validate the code by running it against test cases in controlled environments to verify that it functions according to your requirements.