> ## Documentation Index
> Fetch the complete documentation index at: https://docs.shouldertap.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# Kinds

> A kind determines the schema an expert's answer is captured into. ShoulderTap ships two built-in kinds and lets consumers register their own.

A **kind** is the type of question being asked. It does two things: it routes proposals back to the consumers that `handles_kinds` it, and it determines the JSON Schema the expert's free-text reply is structured into. The canonical schemas live in `spec/schemas/kinds/*.json` in the repository.

## Built-in kinds

<ResponseField name="glossary.definition">
  For defining a term. Structured shape:

  <Expandable title="schema">
    <ResponseField name="definition" type="string" required />

    <ResponseField name="caveats" type="array" required>
      Edge cases or qualifications on the definition.
    </ResponseField>

    <ResponseField name="examples" type="array">
      Optional illustrative examples.
    </ResponseField>
  </Expandable>
</ResponseField>

```json theme={null}
{
  "definition": "A paying account with activity in the last 90 days",
  "caveats": ["Excludes trial accounts", "Q2 reporting uses calendar quarters"],
  "examples": ["An account that logged in and has an active subscription"]
}
```

<ResponseField name="freeform.answer">
  For any other question. Structured shape:

  <Expandable title="schema">
    <ResponseField name="summary" type="string" required>
      A concise answer.
    </ResponseField>

    <ResponseField name="details" type="string">
      Optional supporting detail.
    </ResponseField>
  </Expandable>
</ResponseField>

```json theme={null}
{
  "summary": "Paying accounts active in the last 90 days",
  "details": "Finance treats a login or a billable event as activity."
}
```

## Structuring behavior

When an expert replies, ShoulderTap (with an LLM configured) extracts the structured form and a confidence score. The raw `answer` is always preserved. If extraction fails or confidence is below 0.3, `structured` is `null` but `answer` still holds the expert's exact words — a consumer should always be prepared to fall back to `answer`. Without an LLM configured, `structured` is always `null` and the raw reply passes through verbatim.

## Custom kinds

A consumer can register its own kind with a custom JSON Schema at registration time, via the `kind_schemas` field on [ConsumerRegistration](/api/contracts). A custom schema for a kind takes precedence over a built-in of the same name, so you can also override how a built-in kind is captured for your consumer.