Skill Test Framework

Test framework for validating Claude Code plugin skills across four layers: structure, trigger accuracy, session behavior, and skill value.

Why AI plugins need different testing

Traditional testing verifies deterministic behavior. AI plugin skills are probabilistic — the same prompt can trigger different skills across runs, routing is inferred not explicit, quality degrades silently, and models improve over time (making skills redundant). This framework addresses all four failure modes.

Four-Layer Testing Model

Layer	What it tests	Speed	Script
L1 Structure	Plugin spec compliance, naming, cross-refs	~0.1s	validate.py
L2 Triggers	Skill description accuracy (precision/recall)	~30s/query	run_trigger_eval.py
L3 Sessions	Multi-turn routing, context, boundaries	2-3 min	session_test.py
L4 Value	Does the skill actually help? (with vs without)	5+ min	compare_skill.py

Quick Start

All scripts live in skills/test-framework/scripts/ and accept a --root flag pointing to the plugin directory being tested (defaults to cwd).

# L1: Validate plugin structure
uv run skills/test-framework/scripts/validate.py --root .

# L1: Validate a specific skill
uv run skills/test-framework/scripts/validate.py --root . skills/my-skill/

# L2: Dry-run trigger eval (validate eval set structure)
uv run skills/test-framework/scripts/run_trigger_eval.py \
    --eval-set tests/evals/triggers/my-skill.json \
    --skill-path skills/my-skill \
    --dry-run

# L2: Run trigger eval against claude
uv run skills/test-framework/scripts/run_trigger_eval.py \
    --eval-set tests/evals/triggers/my-skill.json \
    --skill-path skills/my-skill \
    --runs-per-query 3

# L3: Run a session scenario
uv run skills/test-framework/scripts/session_test.py \
    --scenario tests/evals/scenarios/my-workflow.json \
    --verbose

# L4: Compare skill value (with vs without)
uv run skills/test-framework/scripts/compare_skill.py \
    --skill my-skill \
    --scenario tests/evals/scenarios/my-workflow.json \
    --runs 3 --verbose

# All layers for one skill
uv run skills/test-framework/scripts/test_skill.py my-skill

# Test inventory
uv run skills/test-framework/scripts/test_skill.py --inventory

L1: Structure Validation

Validates .claude-plugin/plugin.json, agents, skills, commands, MCP config, and cross-references.

What it checks:

• .claude-plugin/plugin.json — required fields, semver, agent path references
• .mcp.json — server configs have command/url (optional)
• CLAUDE.md — exists with meaningful content (optional but recommended)
• Agents — frontmatter has name + description, names are lowercase
• Skills — SKILL.md exists, name matches directory, description 20-1024 chars
• Commands — files are non-empty, have description in frontmatter
• Cross-references — agents in plugin.json exist, no name collisions
• Orphans — skill directories without SKILL.md

uv run skills/test-framework/scripts/validate.py --root . --json

L2: Trigger Accuracy

Tests whether a skill's description causes Claude to activate for the right prompts. See eval-schemas.md for JSON format.

Creating trigger evals

Create tests/evals/triggers/{skill-name}.json:

{
  "skill_name": "my-skill",
  "evals": [
    {"query": "realistic prompt that should trigger this skill", "should_trigger": true},
    {"query": "near-miss prompt that should NOT trigger", "should_trigger": false}
  ]
}

Guidelines:

• 8-10 should-trigger queries (different phrasings, edge cases)
• 8-10 should-NOT-trigger queries (near-misses, adjacent domains)
• Queries must be realistic and specific (min 10 chars)

Interpreting results

Metric	Meaning
Precision	When the skill triggers, how often is it correct?
Recall	When the skill should trigger, how often does it?
Accuracy	Overall correct rate

Low recall = description too narrow. Low precision = description too broad.

L3: Session Scenarios

Tests multi-turn context, routing accuracy, and skill boundaries.

Creating scenarios

Create tests/evals/scenarios/{name}-workflow.json:

{
  "name": "my-skill-workflow",
  "description": "Test my-skill routing and context",
  "ready_pattern": "❯|\\$|>",
  "steps": [
    {
      "name": "basic-query",
      "prompt": "what does my-skill handle?",
      "timeout": 90,
      "pause_after": 3,
      "assertions": [
        {"pattern": "expected-keyword", "type": "contains", "description": "Routes correctly"},
        {"pattern": "wrong-skill-keyword", "type": "not_contains", "description": "Does NOT invoke wrong skill"}
      ]
    }
  ]
}

Assertion types: contains, regex, not_contains

Safety: Prompts must be read-only. Action verbs (create, delete, push, deploy) are blocked automatically when running with --dangerously-skip-permissions.

L4: Skill Value Comparison

Measures whether a skill actually helps by running the same scenario with and without the skill loaded.

Verdicts

Verdict	Delta	Action
VALUABLE	>+10%	Keep the skill
MARGINAL	+1-10%	Review if context cost is worth it
REDUNDANT	~0% (both high)	Model already knows this — consider removing
INEFFECTIVE	~0% (both low)	Rewrite — skill isn't helping
HARMFUL	Negative	Remove or rewrite — skill makes things worse

When to run comparisons

• After editing a skill's instructions
• When upgrading the underlying model
• Quarterly to prune redundant skills
• Before adding a new skill (baseline first)

Using the Makefile

Copy the Makefile to your plugin root or use ROOT to point at your plugin:

make test                              # L1 + L2 (fast)
make lint ROOT=/path/to/plugin         # L1 only
make integration S=my-skill            # L3 for one skill
make benchmark S=my-skill              # L4 for one skill
make report ROOT=/path/to/plugin       # Test inventory
make test-skill S=my-skill             # All layers

Adding Tests for a New Skill

1. Create the skill (skills/{name}/SKILL.md)
2. Run validate.py to check structure
3. Create trigger evals (tests/evals/triggers/{name}.json) — 8+ positive, 8+ negative
4. Run trigger eval dry-run to validate the eval set
5. Create a session scenario (tests/evals/scenarios/{name}-workflow.json)
6. Run test_skill.py {name} to verify all layers

Reference

• eval-schemas.md — JSON schemas for trigger evals, scenarios, benchmarks