Proactive Trigger Vocabulary

Problem

Agents with many skills face a routing problem: given a user's natural language input, which skill should handle it? Solutions like embedding-based similarity or LLM classification work but are opaque—users don't know what phrases will activate which capabilities.

Additionally, agents may have skills that should activate proactively (without explicit request) when certain topics arise, but without explicit trigger lists, the agent may miss opportunities or activate inappropriately.

Solution

Define an explicit trigger vocabulary for each skill: a list of phrases, keywords, and patterns that should activate that skill. Document these triggers visibly so both humans and agents know the activation criteria.

# Skill definition with explicit triggers
skill: priority-report
description: Generate prioritized task report

triggers:
  exact: ["sup", "priority report", "standup prep"]
  contains: ["what should I work on", "what's pending", "my tasks"]
  patterns: ["what.*on my plate", "action items"]

proactive: true  # Activate without explicit request when triggers match

graph TD A[User Input] --> B{Match Triggers?} B -->|"sup"| C[priority-report skill] B -->|"search hn"| D[hn-search skill] B -->|"check servers"| E[salt-monitoring skill] B -->|No match| F[General response]

Key components:

Trigger lists: Explicit phrases per skill, documented in skill definitions
Proactive flag: Whether skill should auto-activate on trigger match
Priority ordering: When multiple skills match, which takes precedence
User visibility: Triggers documented so users learn the vocabulary

Proactive activation categories:

Information offering: Providing relevant information unprompted
Suggestion: Recommending actions or content
Clarification: Asking for missing information
Correction: Fixing user errors or misunderstandings

User acceptance of proactive activation depends on relevance (contextually appropriate), timing (not interrupting flow), and transparency (explaining why the action was taken).

How to use it

Skill documentation format:

Trade-offs

Pros:

Transparent: Users can learn trigger phrases, feel in control
Predictable: Same input always routes to same skill
Debuggable: Easy to see why a skill activated (or didn't)
Fast: String matching faster than embedding lookup
Documentable: Triggers become part of user-facing docs
Proactive: Agent can jump in when relevant topics arise

Cons:

Rigid: Misses paraphrases not in trigger list
Maintenance: Must update triggers as vocabulary evolves
Conflicts: Multiple skills may want same triggers
Cultural/language bias: Triggers may not translate
Discovery: Users must learn the vocabulary (or read docs)

Hybrid approach:

Combine explicit triggers with semantic fallback:

Check explicit trigger matches first (fast, predictable)
If no match, use embedding similarity (flexible, slower)
Log unmatched inputs to discover new trigger candidates

This hybrid approach (exact match before semantic) is an industry best practice across chatbot platforms, combining predictability with flexibility.

References

Claude Code CLAUDE.md skill documentation pattern
Intent classification in conversational AI
Chatbot trigger/response pattern matching
Slack workflow triggers

Pradhan et al. "Proactive Behavior in Conversational AI: A Survey." ACL 2022
Yang, Shuo, et al. "Should I Interrupt? Proactive Assistance in Human-AI Collaboration." CHI 2021
Primary source: https://github.com/anthropics/claude-code