Intent Detection Prompt for Routing Logic

The raw, unmodified text from the end-user that needs to be classified.

"I can't log into my account and I've tried resetting my password three times."

Required. Must be a non-empty string. Check for PII before logging. If null or empty, the prompt should be aborted and a default fallback intent returned.

A structured list of all valid intents, their descriptions, and their routing targets. This is the model's only allowed output vocabulary.

{"intents": [{"id": "account_access", "description": "User cannot log in or access their account.", "route": "auth_support_queue"}, {"id": "billing_dispute", "description": "User questions a charge or invoice.", "route": "billing_queue"}]}

Required. Must be a valid JSON array of intent objects. Each object must have a unique 'id', a non-empty 'description', and a 'route' string. Validate against the schema before prompt assembly.

The specific intent ID to return when the user input does not match any intent in the catalog.

"general_inquiry"

Required. Must be a string that exists as an 'id' within the [INTENT_CATALOG]. This prevents the model from hallucinating new intents for unknown inputs.

A numeric value between 0 and 1. If the model's self-assessed confidence is below this value, the output should be routed to a human review queue.

0.75

Required. Must be a float between 0.0 and 1.0. If not provided, default to 0.80. A value of 0.0 disables the threshold check and routes all outputs automatically.

A set of 3-5 representative user inputs paired with their correct intent IDs to guide the model's classification behavior.

[{"input": "Where is my refund?", "intent": "refund_status"}, {"input": "Do you have an app?", "intent": "product_inquiry"}]

Optional but strongly recommended. Must be a valid JSON array of objects, each with 'input' and 'intent' fields. The 'intent' value must exist in the [INTENT_CATALOG]. If null, the prompt uses zero-shot classification.

The exact JSON schema the model must use for its response, including fields for the selected intent and a confidence score.

{"type": "object", "properties": {"intent_id": {"type": "string"}, "confidence": {"type": "number"}}, "required": ["intent_id", "confidence"]}

Required. The model's output must be parseable and valid against this schema. A post-generation validation step must check that 'intent_id' is a string from the catalog and 'confidence' is a number between 0 and 1.

A set of 2-3 user inputs that could map to multiple intents, paired with the correct single intent and a brief justification.

[{"input": "I need to change my plan", "intent": "plan_change", "justification": "User explicitly states a desire to change, not just inquire about, their plan."}]

Optional. Must be a valid JSON array of objects with 'input', 'intent', and 'justification' fields. Used to resolve boundary cases. If null, the model relies only on intent descriptions.

string from [INTENT_CATALOG]

Must exactly match a key in the provided intent catalog. Reject on partial match or hallucinated value.

number (0.0 to 1.0)

Must be a float between 0 and 1 inclusive. Reject if confidence < [CONFIDENCE_THRESHOLD] and route to [FALLBACK_QUEUE].

string from [ROUTING_TABLE]

Must map to a valid target in the routing table. Reject if target is not in the allowed set.

string

Must be non-empty and contain a concise justification. Reject if length < 10 characters or if it contradicts the selected intent_id.

boolean

Must be true if no intent matches, false otherwise. Reject if out_of_scope is true but a specific intent_id is provided.

array of {intent_id: string, confidence: number}

If present, each object must pass the same intent_id and confidence checks as the primary fields. Null or empty array is acceptable.

boolean

Must be true if the input is ambiguous across multiple high-confidence intents. If true, the system should not route and instead trigger a clarification prompt.

Output intent is exactly one label from the [INTENT_CATALOG] enum. No extra text, no fabricated labels.

Model returns a label not in the catalog, returns multiple labels when single-label is required, or adds explanatory text.

Schema validation: parse output, check membership in allowed enum set. Fail on any mismatch.

Confidence score is a float between 0.0 and 1.0. Score >= [CONFIDENCE_THRESHOLD] correlates with correct intent assignment.

Score is missing, not a number, outside 0-1 range, or high-confidence predictions are frequently wrong on a golden set.

Run 50 labeled examples through the prompt. Check that precision for predictions above threshold meets [TARGET_PRECISION].

Inputs with no matching intent return [OUT_OF_SCOPE_LABEL] with confidence below [CONFIDENCE_THRESHOLD].

Out-of-scope input is forced into an in-scope intent with high confidence, or valid in-scope input is rejected.

Curate 20 out-of-scope examples. Assert 90%+ receive [OUT_OF_SCOPE_LABEL] and confidence < [CONFIDENCE_THRESHOLD].

When input matches multiple intents, the model selects the single most specific intent per the [INTENT_HIERARCHY] rules.

Model oscillates between two intents on similar inputs, or consistently picks a generic parent intent over a specific child intent.

Create 10 ambiguous input pairs. Check that selection is consistent and follows the documented hierarchy. Measure agreement with human labeler.

The [ROUTING_TARGET] field maps correctly to the selected intent per the [ROUTING_TABLE] mapping.

Routing target is missing, hallucinated, or does not match the mapping defined for the selected intent.

Cross-reference output intent and routing target against the [ROUTING_TABLE]. Fail on any mismatch. This is a hard validation rule.

Output is valid JSON matching the [OUTPUT_SCHEMA] exactly. All required fields present, no extra fields.

Output is not parseable JSON, missing required fields, or contains extra keys that downstream parsers will reject.

Parse output with a JSON schema validator using the exact [OUTPUT_SCHEMA]. Fail on any validation error. Retry once on failure.

When confidence is below [CONFIDENCE_THRESHOLD], the output includes [CLARIFICATION_NEEDED] flag set to true and a suggested clarification question.

Low-confidence prediction is returned without flag, or clarification question is generic and unhelpful.

Test 15 edge-case inputs. Assert that all outputs with confidence < [CONFIDENCE_THRESHOLD] have [CLARIFICATION_NEEDED] = true and a non-empty question.

Prompt injection attempts, gibberish, and extremely long inputs do not produce a valid intent label or crash the output format.

Injection string is classified as a valid intent, or output JSON is malformed due to input length or special characters.

Run a red-team set of 10 injection strings and 5 extreme-length inputs. Assert all return [OUT_OF_SCOPE_LABEL] or a valid error envelope.