Intent Overlap Detection Prompt

This prompt is for AI platform engineers and product teams managing large, production-grade intent taxonomies where a single user input can reasonably map to multiple categories. The job-to-be-done is not primary classification; it is overlap detection. You use this prompt when you need to identify boundary cases between similar intents, quantify how severely an input spans multiple categories, and produce structured overlap data that can drive downstream decisions—such as triggering a clarification flow, routing to a generalist handler, or flagging taxonomy design issues. The ideal user is someone who already has a working classifier and a defined intent catalog, and now needs to harden the system against ambiguous, compound, or edge-case inputs that degrade routing accuracy and user experience.

Do not use this prompt as a replacement for a primary intent classifier. It is a diagnostic and guardrail tool, not a dispatcher. It assumes you already have a list of candidate intents and a user input; its job is to detect when that input sits in the overlap zone between two or more intents. This is most valuable when your taxonomy has fine-grained categories with subtle distinctions—for example, 'refund_request' vs 'order_cancellation' vs 'payment_dispute'—where misclassification is both likely and costly. The prompt works best when paired with a confidence-scoring system: a low-confidence single-label classification can trigger this overlap detector to explain why the input was ambiguous, rather than silently guessing. Avoid using this prompt on extremely short or vague inputs where overlap is inevitable due to missing information; in those cases, a clarification question generator is a better first step.

Before integrating this prompt into a production pipeline, define what you will do with the overlap signal. Common actions include: routing to a multi-intent handler, queuing for human review, logging for taxonomy refinement, or triggering a disambiguation dialogue. The prompt outputs structured overlap pairs with severity scores, which makes it straightforward to set thresholds—for example, only escalate when overlap severity exceeds 0.7. Start by running this prompt against a labeled dataset of known ambiguous cases to calibrate your severity thresholds and identify false positives where the model sees overlap that human reviewers would not. The next section provides the copy-ready template and explains how to adapt the placeholders for your taxonomy.

This prompt template is designed to be the core instruction set for an Intent Overlap Detection system. It takes a user input and a defined taxonomy of intents, then identifies any pairs of intents that could reasonably apply. The output is a structured list of overlapping pairs, each with a confidence score and a brief justification. Use this as the foundation for your model call, wrapping it with your application's input validation, output parsing, and routing logic.

textCopy
You are an intent classification auditor. Your task is to analyze a user's input against a provided list of possible intents and identify any cases where the input could reasonably match more than one intent simultaneously.

## INPUT
User Input: [USER_INPUT]

Intent Taxonomy (Intent Name: Description):
[INTENT_TAXONOMY]

## INSTRUCTIONS
1. Review the user input and the full intent taxonomy.
2. Identify all pairs of intents from the taxonomy that could both apply to this input. An overlap exists if a reasonable person could interpret the input as matching both intents, even if one interpretation is stronger.
3. For each overlapping pair, assign a confidence score from 0.0 (no real overlap, just superficial keyword match) to 1.0 (the input is a textbook example of both intents).
4. Provide a single-sentence justification for each overlap, citing the specific parts of the input that create the ambiguity.
5. If no overlaps are found, return an empty list.

## CONSTRAINTS
- Only consider intents explicitly listed in the provided taxonomy. Do not invent new intents.
- An overlap must involve exactly two distinct intents. Do not return single intents or groups of three or more.
- A confidence score below [CONFIDENCE_THRESHOLD] should not be included in the output.
- Base your analysis strictly on the provided user input and intent descriptions. Do not hallucinate additional user context.

## OUTPUT SCHEMA
Return a valid JSON object with the following structure:
{
  "overlapping_intents": [
    {
      "intent_a": "string (exact name from taxonomy)",
      "intent_b": "string (exact name from taxonomy)",
      "confidence": 0.0,
      "justification": "string"
    }
  ]
}

To adapt this template for your system, start by replacing the [INTENT_TAXONOMY] placeholder with your actual list of intent names and descriptions. The format should be a simple list like - intent_name: description for easy parsing by the model. Set [CONFIDENCE_THRESHOLD] to a value like 0.3 or 0.5 to filter out noise; this threshold should be tuned against a labeled evaluation set. The [USER_INPUT] placeholder is where your application will inject the raw text to be analyzed. Before putting this into production, run it against a golden dataset of inputs with known overlap patterns to calibrate the confidence threshold and ensure the justifications are grounded in the input text, not the model's general knowledge.

The Intent Overlap Detection prompt is designed to sit between your primary intent classifier and your routing dispatcher. It should be invoked when the top two or three predicted intents have confidence scores within a configurable margin (e.g., within 0.15 of each other) or when the top intent falls below a minimum confidence threshold. This prompt is not a replacement for your primary classifier; it is a secondary analysis step that prevents silent misroutes caused by overlapping intent definitions in your taxonomy. The prompt expects a list of candidate intents with their definitions and the original user input. It returns overlapping intent pairs, each with an overlap severity score and a brief explanation of the confusion boundary.

Wire this prompt as a post-classification hook in your routing middleware. After your primary classifier returns top-k intents with confidence scores, check whether the gap between the first and second (or second and third) scores falls below your overlap threshold. If it does, call this prompt with the input and the candidate intent definitions. Validate the output against a strict JSON schema that requires overlapping_pairs as an array of objects with intent_a, intent_b, overlap_severity (a float between 0.0 and 1.0), and boundary_description (a string). Reject any response that does not parse or that contains intents not present in the input candidate list. Log every invocation with the input, candidate intents, raw response, and parsed overlap pairs for audit and taxonomy improvement. If the model returns a malformed response, retry once with a stricter output constraint appended to the prompt. If the retry also fails, escalate the input to a human review queue with the original classification results and a flag indicating overlap detection failure.

Choose a model with strong structured output capabilities for this prompt. The task requires precise comparison of intent definitions and calibrated severity scoring, which smaller or less capable models may handle inconsistently. For high-throughput systems, consider caching overlap results for common intent pairs if your taxonomy is stable—the same pair of intents will produce similar overlap assessments across different inputs. Do not use this prompt for taxonomies with more than 10 candidate intents per invocation; the comparison complexity grows quadratically and the model's attention will degrade. Instead, pre-filter to the top 3-5 candidates before calling overlap detection. If your taxonomy changes, re-evaluate historical overlap logs to identify newly introduced ambiguities before they cause production misroutes.