intent-classifier

Intent Classifier

Classify user requests and route them to the most appropriate skill, command, or specialized agent.

When to Use

• User asks which skill to use or how to route a request
• Need to determine which skill handles a given task
• Multiple skills could handle a request and disambiguation is needed
• Even if they just say "which skill should I use" or "route this"

Classification Workflow

1. Intent Extraction

• Parse the user's natural language request
• Identify primary action (create, update, delete, analyze, etc.)
• Extract entities (file paths, technologies, specific terms)

2. Skill Matching

• Compare against available skills catalog
• Score match confidence for each candidate
• Consider skill descriptions and keywords

3. Route Decision

• Select highest-confidence match
• Handle ties and ambiguous cases
• Fall back to general assistance if no match

Classification Rules

Priority Order

1. Exact keyword match - Skill description contains exact phrase
2. Domain match - Technology stack alignment
3. Action match - Verb/action alignment
4. Context match - File paths, extensions, surrounding context

Confidence Scoring

High (0.8-1.0):   Exact description match, multiple keyword hits
Medium (0.5-0.7): Partial match, related domain
Low (0.3-0.4):    Weak match, shared keywords only
None (<0.3):      No relevant match

Multi-Intent Handling

When a request contains multiple intents:

1. Sequential - Chain skills in dependency order
2. Parallel - Execute independent skills simultaneously
3. Composite - Use parent skill that orchestrates sub-skills

Fallback Strategies

Confidence	Action
> 0.7	Execute primary skill
0.5 - 0.7	Execute skill with confirmation
0.3 - 0.5	Present top 2-3 options to user
< 0.3	Use general assistance

Dynamic Catalog

Update the skill catalog using:

./scripts/dynamic-catalog.sh

This scans .agents/skills/ and regenerates the skill registry with:

• Skill names and descriptions
• Keywords extracted from descriptions
• Compatibility requirements

Skill Selection Examples

User Request	Primary Intent	Routed Skill
"Create a new skill for Docker"	Skill creation	skill-creator
"Audit my code for security issues"	Security audit	security-code-auditor
"Design a REST API for users"	API design	api-design-first
"Fix this bug in my Python script"	Code debugging	(general)

See Also

• skill-creator — Create and improve skills
• skill-evaluator — Evaluate and score skills
• goap-agent — Orchestrator for complex multi-step tasks

Rationalizations

Rationalization	Reality
"I can just pick a skill manually"	Systematic classification ensures the best-matched skill is used and prevents overlooking specialized alternatives.
"Close enough is fine, I'll just pick the first match"	Low-confidence matches lead to wrong workflows and wasted tokens; present options when uncertain.

Red Flags

• Routing to a skill without checking confidence score
• Ignoring multi-intent requests that require chaining skills
• Failing to update the dynamic catalog after adding new skills

References

• references/classification-rules.md - Detailed classification algorithm
• references/skill-catalog.md - Auto-generated skill registry
• references/workflow-patterns.md - Multi-skill orchestration patterns

Voice & Context

• Default: professional + blog
• Reference: voice-profiles skill for definitions and auto-detection.