/llm-classifier

name: llm-classifier
description: LLM-based zero-shot and few-shot classification for flexible intent detection
allowed-tools:
  - Read
  - Write
  - Edit
  - Bash
  - Glob
  - Grep
graph:
  domains: [domain:software-engineering]
  specializations: [specialization:ai-agents-conversational]
  skillAreas: [skill-area:natural-language-processing, skill-area:prompt-instruction-tuning-agents]
  roles: [role:ml-engineer, role:backend-engineer]
  workflows: [workflow:ml-model-lifecycle, workflow:feature-development]

LLM Classifier Skill

Capabilities

• Implement zero-shot classification with LLMs
• Design few-shot classification prompts
• Configure structured output for labels
• Implement confidence scoring
• Design classification taxonomies
• Handle multi-label classification

Target Processes

• intent-classification-system
• dialogue-flow-design

Implementation Details

Classification Patterns

1. **Zero-Shot**: No examples, description-based 2. **Few-Shot**: Example-based classification 3. **Structured Output**: JSON schema for labels 4. **Chain-of-Thought**: Reasoning before classification 5. **Ensemble**: Multiple prompts/models

Configuration Options

• LLM model selection
• Label descriptions
• Example selection strategy
• Output format specification
• Confidence calibration

Best Practices

• Clear label descriptions
• Representative examples
• Consistent output format
• Calibrate confidence scores
• Test with edge cases

Dependencies

• langchain-core
• LLM provider