Intent Router

Intent Router Skill

Use this skill to classify the user's current request into a stable routing result. The model performs the semantic judgment; the route_intent builtin tool validates and normalizes the structured result so downstream agent logic can consume it reliably.

Workflow

1. Read taxonomy.md before choosing task_type, intent_id, recommended_action, or target skill/tool names.
2. Inspect the current user message, recent conversation context, uploaded file metadata, and any known available skills, workflows, databases, or knowledge bases.
3. Decide whether the request can be routed directly or whether required information is missing.
4. Read routing-rules.md when the request could map to a skill, tool, workflow, database, or knowledge retrieval path.
5. Read clarification-rules.md when the request is ambiguous, incomplete, risky, or would trigger database mutation or workflow execution.
6. Read payload-examples.md when building a new or unfamiliar route_intent payload.
7. Call route_intent with a complete structured classification result.
8. Return or use the normalized tool result. Do not invent a downstream tool result; this skill only classifies and recommends next action.

Required Judgments

Always decide:

• task_type: one value from the taxonomy.
• intent_id: stable dotted identifier in the form <task_type>.<subtype> when a subtype is known.
• confidence: numeric value from 0 to 1.
• recommended_action: one value from the action taxonomy.
• normalized_request: concise restatement of what the user is actually asking.
• evidence: short facts from the message/context that support the classification.
• missing_info: required fields that block reliable execution, if any.

Clarification Behavior

This skill usually records ambiguity in missing_info; it does not automatically ask the user.

Use request_user_input instead of route_intent only when a clarification is necessary before a reliable route can be produced, or when the next action would be high impact and the route is not clear enough. Each question must be structured, concrete, and limited to the blocking decision.

References

Read these files as needed:

Constraints

• Do not call downstream business tools from this skill. It only classifies and recommends.
• Do not silently convert an ambiguous request into a specific execution route.
• Do not mark recommended_action as run_workflow, query_database, or call_tool unless the required target is known or listed in missing_info.
• Do not use confidence above 0.85 when multiple materially different task types remain plausible.
• Do not include sensitive raw file contents in evidence; use filenames, MIME types, summaries, or attachment roles.
• Keep evidence short and grounded in the user's message or supplied context.
• Keep output field names stable. Use the route_intent tool result as the source of truth.