wicked-garden:classify

LLM prompt classification analyzes the user prompt using model-driven reasoning to select matching work-shape archetypes from a catalog. It identifies routing signals like blast_radius, novelty, and reversibility, then persists these classifications to SessionState to steer downstream task execution.

You should re-classify a prompt during an active session whenever the scope changes significantly or when a specific classify-due directive is emitted. Mid-session re-classification updates the SessionState so downstream routing adapts to the new task requirements and signals.

To replace regex pattern matching with model-driven reasoning for prompt routing, invoke an LLM-based classifier that reads the prompt and selects the correct work-shape archetypes from a predefined catalog. This approach uses the model's own reasoning to identify signals and persist results to SessionState.

Yes, prompt classification persists its archetype selections and identified signals directly to SessionState. This ensures consistent task routing across multiple turns and prevents drift during mid-session re-classification events.

During prompt classification, the model extracts signals including blast_radius, novelty, reversibility, ambiguity, and scope. These signals are identified through model-driven reasoning and stored in SessionState to guide the selection and execution of downstream playbooks.