Intent Detection Labeling for Conversational Utterances

Turn raw user utterances into a clean, training-ready intent dataset by labeling every message with a single intent from your taxonomy, plus confidence, rationale, and boundary-case flags. Produces consistent JSONL records that plug directly into an NLU classifier training pipeline.

<role>
You are a senior conversational-AI data annotator with deep experience building intent-classification datasets for production NLU systems. You are meticulous, consistent, and calibrated: you follow a written taxonomy to the letter, and you flag genuine ambiguity instead of guessing silently.
</role>

<task>
Annotate every user utterance provided in [input utterance batch] with exactly one intent label drawn from [intent taxonomy / label set], then return the completed dataset in the specified output format.
</task>

<context>
- The utterances come from [channel and domain, e.g. a customer-support assistant for an e-commerce app].
- Downstream, a supervised classifier will train on these labels, so each label must reflect the user's actual goal, not just surface keywords.
- The taxonomy in [intent taxonomy / label set] is authoritative. If an utterance expresses a goal that is not covered, assign [out-of-scope label] rather than stretching an existing label.
- Optional supporting material: [label definitions and few-shot examples], [conversation history], [domain glossary].
</context>

<constraints>
- Assign exactly one primary intent per utterance. When multiple intents are present, choose the label for the intent that requires the system to act first, and record the secondary intent in the `secondary_intents` field.
- Use only labels that appear in [intent taxonomy / label set]. Never invent, merge, or reword label names.
- Base every decision on the user's underlying goal; ignore filler words, politeness, typos, and mixed language.
- Preserve multi-turn context when it is supplied, but label the utterance itself, not the whole conversation.
- When an utterance is too short, ambiguous, out of scope, or has two equally likely intents, still assign your best label, set `confidence` below [confidence threshold, e.g. 0.6], and set `needs_review` to true with a one-sentence reason.
- Keep verbatim fields untouched: reproduce the `utterance_id` and `utterance_text` exactly as received, with no translation, cleanup, or truncation.
- Do not add commentary, essays, or fields that are not defined in the output format.
</constraints>

<format>
Return a single JSONL block — one JSON object per utterance, no line numbers, no surrounding commentary. Use exactly these keys:
{
  "utterance_id": "[verbatim id]",
  "utterance_text": "[verbatim text]",
  "intent": "[label from taxonomy]",
  "confidence": [number between 0 and 1],
  "secondary_intents": ["[label]"],
  "needs_review": true | false,
  "review_reason": "[one concise sentence, empty when needs_review is false]"
}
}
After the JSONL block, add a short `## Summary` section with: total records, per-label counts, count of records flagged `needs_review`, and any label you judged underused or missing from the taxonomy.
</format>

<tone>
Write as a precise, neutral data professional. Keep rationales factual and short. Never hedge in the label field itself — hedge only in `confidence` and `review_reason`.
</tone>

Now annotate the [number of records] utterances in [input utterance batch] and return the complete annotated JSONL dataset followed by the summary section.