Create a classification campaign with answer buttons

When classification is a useful format

Use classification when the result should be a label from a defined set: a support topic, content category, intent or sentiment. The contributor reads an item and chooses an allowed answer, instead of typing a new category each time.

Use a free-text survey when you need an explanation in the person’s own words. A button is convenient only if the available choices can represent the answer. Tellseed supports classification labels as buttons, so contributors do not need to type the category name.

Define labels that different people can apply

Write a short definition for every category and include an example near the boundary between similar categories. Decide how to handle ambiguous items, missing context and examples that fit none of the labels.

For a support triage task, Billing, Technical problem, Feature request and Other can be a starting point. Test those categories against actual examples before sending the campaign to a larger group. Remove personal information that is unnecessary for the judgement.

Ask your agent to build the classification task

Connect your workspace to an agent, then describe the item pool and allowed labels. State how many items you want to prepare and how many each person should see. Those are two different settings.

Choose the item pool and the round length

For a small pilot, prepare ten examples and show all ten to each contributor. For a larger task, you can ask for a pool of one hundred examples with ten selected per round. Have the agent report both settings so a large pool does not accidentally become a long questionnaire.

If you need several independent labels per item, explicitly define that requirement and verify coverage in the collected data. A short round does not by itself guarantee balanced coverage or agreement between reviewers.

Check the preview for real buttons, exact category names, item content and progress. Avoid using a generated example pool as if it were evidence about real customers.

Inspect the labels before using the dataset

Approve the draft, publish and share the contributor link with the people doing the labelling. Ask your agent for the private answers view and an export containing item identifiers and chosen labels.

Check for missing categories, skipped items and examples that repeatedly confuse people. If several people label the same item, examine disagreements before collapsing their labels into a single answer.

For training or evaluation data, document the instructions and the campaign version alongside the export. Human labels need a clear interpretation; collecting them does not automatically make them a reliable benchmark.