What LLMs are good at
Classification
Classification asks a model to choose among labels you defined, instead of writing an open-ended answer.
Mental model
You hand the model a menu. It points at one item, and it can say when nothing on the menu fits.
Labels route work. A ticket, a document, or a user message becomes something your code can switch on.
When to use
- The set of outcomes is known and small enough to list.
- A wrong label has a clear cost, so you can sample and score it.
When not to use
- The categories change every day and nobody can name them.
- You actually need the underlying fields, not a single bucket.
Example
Labels for a support message: billing, bug, how_to, cancel, other.
"I was charged twice and I want last month refunded" is billing. It is not cancel just because the customer is unhappy. The label list is the product decision. The model only picks.
Add other so the model has an honest way out. Then read what landed in other. That is how you discover a missing label.
Implementation notes
Write each label with a one-line rule and one example that would be easy to confuse with a neighbor. "Billing means money movement. Cancel means ending the account, even if a refund is mentioned."
Return the label plus a short evidence quote from the input. The quote makes reviews faster and makes fabricated confidence obvious.
If two labels are often both right, change the task. Multi-label classification, or extraction of reasons, is clearer than forcing one bucket.
Common mistakes
- Hiding the label rules and hoping the names are self-explanatory.
- Using dozens of overlapping labels on the first attempt.
- Treating the model's own "confidence" number as a probability you can threshold blindly.