Better results with AI

Getting consistent, high-quality output when using the FirstSpirit AI Suite requires more than activating the feature. While pre-set prompts are available for each assistant, outputs become even better when prompts are customized. This page covers the practices that make the difference: how to write prompts that produce tailored results, how to use the right assistant for each task, and how to review output responsibly.

The assistants cover different scopes. Selecting the right one for a task reduces unnecessary steps and produces more relevant output. For a list of which assistant is best suited for a given task see What the assistants do.

The output an AI model produces is directly dependent on the prompt it receives. A vague prompt produces a generic result. A specific, well-structured prompt produces output that is closer to what you need.

Consider the following guidelines when writing prompts:

• Be specific and concise: State the task, the context, and the expected output format explicitly. A longer prompt is not necessarily a better one. Include only the information the model needs to complete the task.
• Define a role: Opening a prompt by assigning the model a role produces more focused output. For example: "You are an expert in writing and text creation in German". The role should be as specific as the task requires.
• Break complex tasks into steps: If a task involves more than one action, list the steps in numbered order and instruct the model to perform them in strict sequence. "Read through the steps and understand them before you begin" is a useful instruction to include.
• Delimit your input: Separate the instruction from the input text clearly. FirstSpirit Assistants insert your selected content automatically at the end of the prompt. You do not need to add the input manually.
• Specify the output: Define exactly what the model should return and nothing else. For example: "Output only the rewritten text. Do not add anything else to your answer." This prevents the model from adding explanations or commentary that you did not ask for.
• Iterate: The first version of a prompt is rarely the final one. Analyze the output, identify where it falls short, and make one change at a time. Minor changes, including formatting and line breaks, can shift the output noticeably. Once a prompt produces consistent results, reuse it.

The effectiveness of any prompt depends on the model connected to your project, your configuration, and the specific task. Results will vary.

The FirstSpirit AI Suite ships with a set of ready-made operations for the Content Assistant, Analyze Assistant, and Image Assistant. These cover common tasks such as shortening text, improving quality, and adjusting tone. They are a reliable starting point.

However, you can achieve more tailored and effective output by using use-case specific operations. For most tasks, customizing an available project operation is more effective than writing a prompt from scratch. Project and brand prompts encode your organization's tone, style rules, and channel requirements. Using them produces consistent output across editors and content types without each editor having to apply those rules manually.

Keep customizations scoped to the project where they are needed. For most use cases, a project-level operation is the appropriate level. This avoids unintended changes to other teams or projects.

You can find customization instructions and examples here.

The assistants are designed to complement each other. Using more than one across a content task produces more complete results than using a single assistant in isolation.

The following is an example workflow for a standard content publishing task:

• Use the Content Assistant to draft or rephrase the text in a field.
• Use the Analyze Assistant to run a page-wide tone, accessibility, and SEO check before release.
• Use the Image Assistant to generate alt text and tags for images on the page in bulk.

Each assistant operates within its defined scope. Every result is a suggestion you can accept or reject.

You always have to review AI output before you release or publish it. When doing so, keep the following in mind:

• Hallucinations: Large Language Models can produce statements that are represented as fact but are incorrect or fabricated. Verify any factual claims in the output, particularly names, figures, or references.
• Bias: Training data can contain biases that are reproduced in the model's output. Review output critically, particularly for language that may be skewed or unrepresentative.
• Knowledge cutoff: Models are trained on data up to a fixed point in time. Output may not reflect recent developments, terminology changes, or current best practices.
• Output quality varies by connected model: The model connected to your project directly affects the quality of the output. If results are consistently below expectations, raise this with your project admin.