Persona Skills for Claude
The Hierarchy of Agentic Reliability: From Routing to Orchestration
Key Finding: Natural Language Routing achieves 81%+ reliability for single-shot tool selection using few-shot examples and negative constraints. But it fails completely on compound, multi-step prompts. We found the architectural ceiling.
What We Did
We used Anthropic's Agent Skills system to shape Claude's behavioral voice/persona rather than its procedural capabilities. This is off-label but effective—skills are designed for task workflows, but work for disposition shaping too.
We ran a 16-phase evaluation testing activation rates, failure patterns, and architectural limits.
The Hypothesis
Can Claude's Skills system be used for persona/voice shaping rather than procedural task completion?
Traditional use: "When user wants to create a PDF, load these instructions."
Our use: "When user wants exploratory dialogue, adopt this conversational voice."
Key Findings
1. Activation is Bimodal
The overall activation rate of 56% is misleading. Activation is actually bimodal:
| Prompt Type | Activation Rate |
|---|---|
| Abstract trigger phrase only | ~0% |
| Concrete dilemma with named options | ~100% |
| Mixed/ambiguous | Variable |
2. Few-Shot Examples Double Activation
Description phrasing matters enormously:
- Semantic definitions only: 39% activation
- With few-shot examples: 81% activation
Examples create a "semantic cone"—the model interpolates between them rather than litigating against rigid definitions.
3. Three Failure Classes (All Solved)
| Failure Class | Symptom | Solution |
|---|---|---|
| Ambiguity | Wrong skill activates | Few-shot examples (semantic cone) |
| Overconfidence | Skill activates when shouldn't | PROHIBITED constraints |
| Collision | Overlapping namespaces | Scope exclusion |
4. The Hard Limit: Compound Prompts Fail
Phase 16 Discovery: Natural Language Routing is single-shot. Compound prompts like "do X, then trigger skill Y" fail due to:
- Token Gravity: Early tasks absorb all attention ("Skip-to-Panel")
- Context Decay: Later instructions fade ("First-Intent-Wins")
Conclusion
Natural Language Routing is sufficient for Tool Selection but insufficient for Task Orchestration.
Multi-step workflows require a dedicated Supervisor state-manager or explicit Human-in-the-Loop intervention.
Recommendations
| Use Case | Recommended Approach |
|---|---|
| Single-shot tool selection | Natural Language Routing |
| Multi-step workflows | Human-in-the-Loop OR Supervisor pattern |
| High-reliability routing | Few-shot examples + PROHIBITED constraints |
| Persona/behavioral skills | Observable signals + verbal tics + examples |
Full Guide
The complete 1,450-line practitioner's guide includes:
- 16-phase experiment narrative
- 48 systematic test results
- Skill templates and examples
- Expert Panel Framework
- Reproducible test harness
Read the full guide on GitHub →
Materials
All materials available at: github.com/credentum/vivarium-lab
- Guide:
persona-skills/guide.md - Skill descriptions:
persona-skills/skill-descriptions-v3.md - Example skill:
persona-skills/SKILL.md
This research was conducted collaboratively by Matt + Claude (Opus 4.5). Study complete December 2025.