athola/subagent-testing

Subagent Testing - TDD for Skills

Test skills with fresh subagent instances to prevent priming bias and validate effectiveness.

When NOT To Use

• Writing the skill under test (use abstract:skill-authoring)
• A static quality audit with no execution (use abstract:skills-eval)

Table of Contents

1. Overview
2. Why Fresh Instances Matter
3. Testing Methodology
4. Quick Start
5. Detailed Testing Guide
6. Success Criteria

Overview

Fresh instances prevent priming: Each test uses a new Claude conversation to verify the skill's impact is measured, not conversation history effects.

Why Fresh Instances Matter

The Priming Problem

Running tests in the same conversation creates bias:

• Prior context influences responses
• Skill effects get mixed with conversation history
• Can't isolate skill's true impact

Fresh Instance Benefits

• Isolation: Each test starts clean
• Reproducibility: Consistent baseline state
• Measurement: Clear before/after comparison
• Validation: Proves skill effectiveness, not priming

Testing Methodology

Three-phase TDD-style approach:

Phase 1: Baseline Testing (RED)

Test without skill to establish baseline behavior.

Phase 2: With-Skill Testing (GREEN)

Test with skill loaded to measure improvements.

Phase 3: Rationalization Testing (REFACTOR)

Test skill's anti-rationalization guardrails.

Quick Start

# 1. Create baseline tests (without skill)
# Use 5 diverse scenarios
# Document full responses

# 2. Create with-skill tests (fresh instances)
# Load skill explicitly
# Use identical prompts
# Compare to baseline

# 3. Create rationalization tests
# Test anti-rationalization patterns
# Verify guardrails work

Detailed Testing Guide

For complete testing patterns, examples, and templates:

• Testing Patterns - Full TDD methodology
• Test Examples - Baseline, with-skill, rationalization tests
• Analysis Templates - Scoring and comparison frameworks

Success Criteria

• Baseline: Document 5+ diverse baseline scenarios
• Improvement: ≥50% improvement in skill-related metrics
• Consistency: Results reproducible across fresh instances
• Rationalization Defense: Guardrails prevent ≥80% of rationalization attempts

See Also

• skill-authoring: Creating effective skills
• bulletproof-skill: Anti-rationalization patterns
• test-skill: Automated skill testing command

Exit Criteria

• [ ] Baseline (RED) phase documents at least 5 diverse scenarios run in fresh Claude instances without the skill active, with full response text recorded.
• [ ] With-skill (GREEN) phase uses identical prompts in new fresh instances (not continuations of the baseline conversation) and shows >= 50% improvement on skill-related metrics.
• [ ] Rationalization (REFACTOR) phase shows skill guardrails blocking >= 80% of rationalization attempts tested across at least 3 pressure scenarios.
• [ ] Results are reproducible: the same prompts in a new fresh instance produce consistent outcomes, confirming the effect is not conversation-history priming.