Optimize Prompt-Based Custom Models
min read - Updated on Oct 28, 2026A prompt-based custom model classifies customer feedback from the names and descriptions of its labels, so the way you write them decides how accurately the model works. This article shows how to write labels and descriptions the AI can tell apart, how to word the question of a Dimension model, how to read a model's consistency score and its categories as signals of label quality, and how to test a model and replace it with a better version. If you haven't created a model yet, start with Create a Custom Model with a Prompt.
Custom models are available in the paid packages, not in Starter (see kimola.com/pricing). Sign in to Kimola and open Models in the main menu: your models are listed there, and Create AI Model › By Prompt starts a new one.
How Your Labels Shape the Results
A prompt-based model doesn't learn from labeled examples. When it classifies a piece of feedback, the AI reads the text against the name and the description of every label and assigns the labels that fit. The labels are the model's whole instruction, so every improvement in this article is an improvement to them.
You write the labels in the Labels step of the prompt wizard, and they stay as they are once the model is created. A model's Settings page changes only its name, and Organize labels changes only how its labels are grouped. To change a label or its description, you create a new version of the model and use it in place of the old one (see "When to Rebuild" below). The best time to apply this article is therefore before you leave the Labels step, and again whenever a test shows that a label doesn't work.
The labels and descriptions of a prompt-based model can't be edited after the model is created. To change them, create a new version of the model.

Refine Your Labels
Each label is one tag the model can assign, and it appears as a theme in Reports and Spaces. The clearer and more distinct your labels are, the more accurately the model sorts feedback between them. Check every label against these points.
One topic per label. A label covers one subject customers talk about, such as Pricing, Delivery Issues or Mobile App. A label that joins two subjects, such as "Pricing and Delivery", catches feedback about either one and tells you little about both.
No overlapping labels. Two labels with nearly the same meaning, such as Customer Support and Customer Service, split the same feedback between them unpredictably. Merge them into one label, or give each a clearly different scope, such as Call Center and Live Chat.
Neutral names for topics. In a content classifier, name the topic rather than an opinion about it: Pricing rather than "Too Expensive", Delivery rather than "Late Delivery". Kimola measures how customers feel about each theme separately, so a neutral label collects both the praise and the complaints about a topic and shows the balance between them. A label with an opinion in its name pulls the AI toward feedback that shares the opinion and leaves the other side of the topic without a label.
Short names. Use a few words that a reader understands without the description, because the name is what your team reads as the theme in every Report and Space.
A focused number of labels. A model needs at least two labels, and for the best accuracy keep to no more than 25. A short list of well-separated labels classifies more consistently than a long list of narrow labels that overlap. If you need many labels, group them into categories so they stay easy to read.
A mobile network operator's first draft has the labels Network Problems, Bad Coverage, Slow Internet and Customer Care Complaints. The first three overlap, and every name carries a complaint. A refined version uses Coverage, Mobile Data Speed, Billing and Customer Care: each covers one topic, and a customer praising a fast connection is counted under Mobile Data Speed next to the customers who complain about it.
Read your list of label names on its own, without the descriptions. If two names could describe the same piece of feedback, the AI will hesitate between them too.
Improve Label Descriptions
Descriptions tell the AI what kind of feedback belongs under each label, and the AI goes by them when a piece of feedback could fit more than one label. The wizard's own example shows the level of detail: a Pricing label described as "Conversations about product cost, discounts, and value for money."
A good description says in one or two sentences what the label covers and names the kinds of feedback that belong to it: the words customers use, the situations they describe and the parts of the product or service they mention. Write it in your customers' terms: customers write "I was charged twice" where your team might say "payment reconciliation error". Where two labels are close, let the descriptions draw the line between them. Avoid a description that only repeats the name, such as "Feedback about pricing", and avoid descriptions that read alike.
A software company's support tickets mention invoices and sign-in problems, sometimes in the same ticket. Descriptions that keep the two labels apart might be: Billing, "Invoices, charges, refunds, plan changes and renewals"; Account Access, "Signing in, passwords, two-factor codes, locked accounts and adding teammates." A ticket that says "I can't sign in to download my invoice" then fits both labels for a clear reason, and with multi-label classification it gets both.
A custom sentiment model always uses the two labels Positive and Negative, so its descriptions are the only thing you shape. The wizard fills them in ("Feedback expressing satisfaction, praise, or a good experience." and "Feedback expressing dissatisfaction, complaints, or a bad experience."), and you can edit them to say what counts as each for your business, for example that a request for a refund counts as negative.
Write a Clear Question for a Dimension Model
A Dimension model answers an analytical question about your feedback, and its Question and Description guide the AI the way label descriptions do. When you choose a Dimension from Kimola's list, its question and description are filled in for you to fine-tune. When you write your own, ask one thing that customers actually answer in what they write, and use the description to say what counts as an answer. The wizard's example is the question "Which part of the journey is the customer talking about?" with the description "Identifies the stage of the customer journey, from discovery to after-sales support, described in the feedback."
The question matters even more with Assign labels based on the dataset at runtime. In that case you define no labels, and the AI discovers the answers from each dataset by answering your question right before classifying it, so the question and its description are all it has. A broad question gives answers that change from one dataset to the next. When the answers you see have settled into a list you want to keep, create a new model with those answers as defined labels: every dataset is then counted under the same answers, and the model can also be added to a Funnel, which takes only Dimension models with defined labels.
A hotel group that asks "What do guests mention?" gets answers that overlap with its content classifier. Asking "Who is the guest traveling with?" and describing it as "The company the guest describes on the stay: family, a partner, colleagues, friends or no one" gives clear answers such as With family, Business trip and Traveling alone.
Check the Consistency Score
Every custom model shows a consistency score: on the Models list next to the model's name, and on the model's Overview page as Consistency, "Percentage of semantically consistent pairs." A high consistency means the model's labels are told apart clearly. A low score is a sign to look again at labels that overlap and at descriptions that are too general or read alike. See What is Model Consistency?
The score is most useful as a comparison. When you create an improved version of a model, compare its consistency with the old version's to see whether your changes made the labels clearer. On the Models list, Least Consistent First brings the models that need attention to the top.

Sort the Models list by Least Consistent First to find the models whose labels to review first.
Check How Your Labels Group into Categories
Categories group the labels of a model under a few topics, so that Reports and Spaces read like a table of contents. Group my labels into categories, in the Categories step of the wizard or under Organize labels on the Models list, drafts the categories from your labels in a few seconds. Your feedback is still classified with the labels, so categories never change the results.
The draft is also a quick reading of your labels. Labels the draft puts side by side in one category may be closer in meaning than you intended, so check that their descriptions keep them apart. A label left under "Labels that fit no category." may be too broad, or about something else than the rest of the model. When the draft groups your labels the way you would, your labels are clear. Unlike the labels, categories can be changed at any time without creating a new model: rename them, describe them, move labels between them or remove them. See Organize Labels into Categories.
An online store's draft puts Returns and Refunds in a category of their own. If the two labels are meant to count different feedback, their descriptions need a clearer line, such as sending the item back for Returns and getting the money back for Refunds. If they aren't, one label is enough.

Test and Evaluate
Before you use a model in an analysis, test it on real feedback in the test console on its Overview page. Under Text, you type or paste your own text and the predictions appear as you type. Under Document, you upload a file in .xls, .xlsx, .csv or .tsv format, smaller than 30 MB. The Multi-Label and Strict Mode options set how the predictions are made. See Test a Custom Model.
Test with feedback your team knows well, including the difficult cases: texts that mention two topics, short texts and texts with mixed feelings. When you review the predictions, check whether similar texts get the same labels, whether some labels are rarely or never used (a sign they are unclear or not needed), and whether two labels are often confused (a sign their descriptions need more contrast).

When to Rebuild
Because the labels of a prompt-based model can't be edited, improving them means creating a new version. Use Create AI Model › By Prompt with your refined labels, descriptions or question, and give the model a new name, such as "Support Topics v2", because every model in your account needs a name of its own. Test the new version on the same texts as the old one and compare their consistency.
Once the new version works better, use it in its place. Choose it in your next Reports. If the old model is in a Funnel, edit the Funnel to use the new one: the files uploaded to its Spaces after the edit are classified with the new model, and records already classified keep their labels. See Manage Funnels. When you no longer need the old version, delete it from its Settings page with Delete Model.
Delete Model deletes the model with all its records, and it can't be undone.
If a prompt-based model keeps mislabeling feedback even with clear labels and descriptions, the distinctions you need may be easier to show with examples than to describe. If your team already has labeled feedback, create a model with it instead. See Create a Custom Model with a Training Dataset and Optimize Dataset-Based Custom Models.