Create a Custom Model with a Prompt
3 mins read - Updated on Nov 12, 2026You can create a custom model with a prompt by telling Kimola what the model is for and describing its labels in plain language, with no training dataset. A wizard in Models takes you through it: choose the purpose (a content classifier, a sentiment classifier or a Dimension), define the labels with a name and a description each, group the labels of a content classifier into categories if you like, and name the model. The model is then ready to classify customer feedback in your Reports and Funnels.
Custom models are available in the paid packages (Basic, Standard, Business and Enterprise), not in the free Starter package. See kimola.com/pricing.
Sign in to Kimola, open Models in the main menu, click Create AI Model and choose By Prompt. If you have no custom model yet, the page offers Create from Prompt instead.

Step 1: Choose the Purpose
The first step is Purpose: "Tell us what your model should do with your customer feedback." Under What will your model be used for? you choose one of two kinds of model.
Classifier "Tags every piece of feedback with the labels you define, such as topics or sentiment." When you choose it, What kind of classifier? asks for one of two kinds. A Content Classifier "Sorts feedback into the topics that matter to your business." A Sentiment Classifier "Detects how customers feel in each piece of feedback."
Dimension "Answers an analytical question about your feedback, such as who is talking or what goes wrong." Where a classifier tags what a piece of feedback is about, a Dimension model asks one research question of every piece of feedback, and its labels are the possible answers.

When you choose Dimension, Which dimension? lists Kimola's 10 Dimensions with their questions, from Motivations ("Why is this chosen or preferred?") and Pain Points ("What goes wrong or creates difficulty?") to Co-Occurrences ("Alongside what does this occur?"). Choosing one fills in its Question and Description: "This is the question your model will answer for each piece of feedback. You can fine-tune it for your business." The last item, Custom Dimension, lets you "Write your own question for the model to answer." with a description of what the model should find.
For an online course platform, a Custom Dimension might ask "What made the learner stop or finish the course?", with the description "Identifies the reasons learners give for dropping out of a course or completing it."
Click Continue to go on. If the question or its description is empty, the wizard asks you to "Please go back and fill in the question and its description."

Dimension models are used in Funnels, which classify the feedback of your Spaces. Funnels are available in the Enterprise package. A Report doesn't take a Dimension model: you choose its Dimensions from Kimola's list while you create it.
Step 2: Define the Labels
The second step is Labels. A prompt-based model learns from no examples, so the name and description of each label are what the AI reads to decide which piece of feedback gets which label. What the step asks for depends on the purpose you chose.
For a content classifier, "Create a list of labels with short descriptions to help the AI classify your text accurately." Click Add a New Label for each label and fill in its Name (a short name, such as Pricing) and Description (a brief explanation, such as "Conversations about product cost, discounts, and value for money."). Add at least two labels to continue. A model holds up to 25 labels, and the counter above the list shows how many you have, for example "4 of 25 labels".
Name the topic, not the opinion about it: Pricing rather than Too Expensive. A piece of feedback about price then gets the same label whether it praises or criticizes, and the sentiment classifier says which.
Write each description so that it tells one label apart from the others. When two descriptions cover the same ground, the AI has no clear reason to choose between them.

For a sentiment classifier, "Describe what positive and negative mean for your business, so the AI judges sentiment the way you do." "Sentiment models always use these two labels. You can edit their descriptions." The labels are Positive ("Feedback expressing satisfaction, praise, or a good experience.") and Negative ("Feedback expressing dissatisfaction, complaints, or a bad experience."). You can't rename them or add others, but you can rewrite their descriptions to say what counts as each for your business.
For a Dimension model, "Define the possible answers to your question as labels with short descriptions." Your question and its description are shown above the list. Add the answers as labels, each with a name and a description, at least two of them.
A Dimension model can also skip this work. Turn on Assign labels based on the dataset at runtime: "Skip defining labels now. The AI will discover them from each dataset by answering your question, right before classifying it." The answers then come from the feedback itself, so a new kind of answer can show up without anyone adding a label for it. Define the labels yourself when the answers have to stay the same from one dataset to the next.
A mobile game studio might build an Unmet Needs model with runtime labels. In one month's reviews the AI might find answers such as more save slots and offline mode, and in the next, after an update, controller support.
The wizard checks the labels before you continue: "Name cannot be empty.", "Name must be unique." and "Description cannot be empty." Click Continue when the list is complete, or Back to change the purpose.
Step 3: Group Labels into Categories (Optional)
A content classifier has one more step, Categories: "Group your labels under a few topics, so that your reports read like a table of contents instead of a long list." Sentiment and Dimension models skip it and go straight to naming. Feedback is still classified with your labels; the categories only change how Reports and Spaces show them, with the labels grouped under their categories.
Categories help when a model has many labels. With fewer than 4 labels, the step says your labels "read well on their own; add at least 4 to group them." and you can click Continue.
With 4 labels or more, Group my labels into categories is on, and Kimola drafts the categories from your labels ("It takes a few seconds."). "We drafted the categories from your labels: rename them, describe them and move labels between them. Your feedback is still classified with your labels." A line above the categories shows how many labels sit in how many categories.

You can then shape the draft. Give each category a name and a description of what customers talk about in it. Move a label to another category, or to Other (no category), from the menu on the label. Add a category adds an empty one (a model can have up to 12 categories), and the delete icon on a category removes it. Suggest again drafts a new set of categories from your labels.
Labels you keep out of every category are listed under Other: "Labels that fit no category. Your reports show them together under Other; drag one into a category to group it."
An online furniture store might have labels such as Delivery Time, Damaged Items, Assembly, Material Quality, Pricing and Customer Support. The draft might group them as Orders and Delivery (Delivery Time, Damaged Items), Product (Assembly, Material Quality) and Service (Pricing, Customer Support).
If you don't want categories, turn Group my labels into categories off and click Continue: the model's labels are then read as a single list. If the draft fails, the step says "Your labels could not be grouped this time." Click Try again, or Group them myself to make the categories by hand.
Every category needs a name, a description and at least one label. If one is missing, the wizard asks you to go back and check your categories before it creates the model.
You don't have to decide now. You can group the labels of a model later, or remove its categories, with Organize labels on the Models list. See Organize Labels into Categories.
Step 4: Name Your Model
The last step is Naming: "Give your model a name you'll recognize when you use it in your reports." Under the name field, the wizard sums up the model: a content classifier with its number of labels, for example, or a Pain Points dimension with its question and "Labels assigned from the dataset at runtime".
Choose a name that says what the model does, such as "Furniture Store Themes" or "Course Feedback Sentiment", because it is how you and your team find the model when you build a Report or a Funnel. Each of your models needs its own name: if the name is taken, the wizard says "You already have a model with this name. Please choose another one."

Click Create Model. The model appears on the Models list with its kind, its number of labels and its date. While Kimola prepares it, the list may show Waiting... or Training... in its place; when that disappears, the model is ready. If the list shows Error, the model couldn't be prepared; create it again.
Use Your Model
A new model works like any other custom model. Open it from the Models list with View to see its consistency, and try it before you use it in an analysis: the console classifies your own text as you type, or a file you upload. See Test a Custom Model.
A content or sentiment classifier can be used in a Report: in the Classifiers step of a new Report, choose it under CUSTOM MODELS, in place of or next to the pretrained models. See Create a Report Using a Custom Model. Whether one piece of feedback can get one label or several is set by the Report, in its Structure step (see Multi-label vs. Single-label Classification).
In the Enterprise package, every kind of prompt-based model can go into a Funnel: one content classifier, one sentiment classifier and one Dimension model per Dimension. Every Space that uses the Funnel classifies each new file with them. See Create a Funnel.
Every model you add to a Report or a Funnel uses queries: queries = rows × models × Dimensions. See How to Calculate Queries.
If the results aren't what you expected, the usual cause is a label whose description overlaps with another or doesn't match how customers talk. See Optimize Prompt-Based Custom Models.