testing-skills-with-subagents
Use after writing a new skill and/or when testing existing skills, creating skill evaluations, or verifying skills work under pressure - applies TDD/RED-GREEN-REFACTOR to skill documentation by running baseline tests, measuring compliance, and closing rationalization loopholes
When & Why to Use This Skill
This Claude skill automates the testing and validation of AI agent skills using a subagent-driven TDD (Test-Driven Development) approach. It ensures skill reliability by applying the Red-Green-Refactor cycle to process documentation, running baseline tests, and measuring compliance to eliminate rationalization loopholes and ensure deployment-ready performance.
Use Cases
- Skill Development Validation: Use the 'Fast' variant during the coding phase to quickly iterate and verify that new skill logic meets documentation requirements.
- Pre-deployment QA: Execute the 'Slow' variant for rigorous, worktree-based isolated testing to gain high confidence before pushing skills to production.
- Stress Testing: Verify if skills maintain performance and logic consistency under high-pressure scenarios or complex edge cases.
- Regression Testing: Run baseline tests on existing skills after updates to ensure new changes haven't introduced rationalization loopholes or compliance failures.
| name | testing-skills-with-subagents |
|---|---|
| description | Use after writing a new skill and/or when testing existing skills, creating skill evaluations, or verifying skills work under pressure - applies TDD/RED-GREEN-REFACTOR to skill documentation by running baseline tests, measuring compliance, and closing rationalization loopholes |
Testing Skills With Subagents
Overview
Testing skills is just TDD applied to process documentation.
Choose your testing approach based on development phase:
- Fast variant: Quick iteration during skill development (15-30 min)
- Slow variant: Rigorous validation before deployment (45-90 min)
Announce at start: "I'm using the testing-skills-with-subagents skill."
Workflow
Follow this workflow exactly - do not skip steps:
graph TD
a@{ shape: stadium, label: "Start: Announce Skill Usage" }
b@{ shape: rect, label: "Ask Variant Question" }
c@{ shape: diam, label: "Variant Choice?" }
d@{ shape: rect, label: "Read variants/fast-conversational.md" }
e@{ shape: rect, label: "Read variants/slow-isolated.md" }
f@{ shape: stadium, label: "Execute Variant Workflow" }
a --> b
b --> c
c -->|Fast| d
c -->|Slow| e
d --> f
e --> f
classDef start fill:#ccffcc
classDef decision fill:#ffffcc
classDef action fill:#ccccff
a:::start
c:::decision
b:::action
d:::action
e:::action
f:::start
Design Rationale: Mermaid flowchart provides visual enforcement of workflow sequence, preventing LLM from skipping announcement or variant question.
Step 1: Choose Testing Variant
Use the AskUserQuestion tool with these exact parameters:Question: "Which testing variant do you want to use?"
Options:
- "Fast: Conversational testing with control scenarios" - 15-30 min iteration, lightweight logging, good for skill development
- "Slow: Worktree-based isolated testing" - 45-90 min validation, full infrastructure, deployment-ready confidence
Trade-offs:
| Aspect | Fast Variant | Slow Variant |
|---|---|---|
| Time | 15-30 min | 45-90 min |
| Infrastructure | Lightweight logs | Full worktree isolation |
| Confidence | Moderate | High (deployment-ready) |
| Best For | Iteration, hypothesis testing | Pre-deployment validation |
Step 2: Execute Selected Variant
Based on the user's variant choice from Step 1:IF user selected "Fast: Conversational testing":
- Read the file: variants/fast-conversational.md
- Follow ALL instructions in that file
IF user selected "Slow: Worktree-based testing":
- Read the file: variants/slow-isolated.md
- Follow ALL instructions in that file
Do NOT proceed without reading the selected variant file.
Error Handling
If variant file doesn't exist: Display this error message:
❌ Error: Variant file not found
Expected file: variants/{variant-name}.md
This indicates incomplete skill installation. Please check:
1. File structure is correct (.claude/skills/testing-skills-with-subagents/variants/)
2. Variant files exist (fast-conversational.md, slow-isolated.md)
3. Repository is up to date
Cannot proceed without variant file.
Design Rationale:
<critical-instruction>tags ensure LLM doesn't skip file read- Context passed forward but doesn't override variant's workflow
- Clear error message for missing files