Evaluation Skills

The eval-* family measures the quality of AI agent outputs, prompts, and orchestration results. Skills use a 3-way majority vote pattern to reduce individual model bias.

Skills

Skill ID	Description	Model Class
eval-prompt	Scores a prompt on clarity, specificity, context sufficiency, and output predictability	free
eval-output-grading	Grades a model response against rubric criteria (accuracy, format, completeness, safety)	free
eval-variance	Measures response stability across multiple runs of the same prompt	cheap
eval-prompt-bench	Benchmarks a prompt against alternatives using the same input; ranks by score	cheap
eval-adversarial	Tests prompt robustness against injection, jailbreak, and edge-case adversarial inputs	strong

Voting Pattern

eval-output-grading
  ├── Claude Haiku 4.5   → vote A
  ├── GPT-5 mini         → vote B
  └── GPT-4.1            → vote C
          │  (split?)
          └── GPT-5.4    → tiebreak
                │  (still split?)
                └── Claude Sonnet 4.6 → final

When to Use

Situation	Skill(s)
Testing a new prompt variant	eval-prompt + eval-variance
Comparing two prompt approaches	eval-prompt-bench
Grading an agent’s response	eval-output-grading
Security review of prompt surface	eval-adversarial

Instructions That Invoke These Skills

• evaluate — all five skills coordinated
• benchmark — uses eval-output-grading + eval-variance for blind comparison
• prompt-engineering — uses eval-prompt to score generated prompts
• govern — uses eval-adversarial for injection resistance testing

Score Format

Skills output a structured score object:

{
  "skill": "eval-output-grading",
  "score": 8.5,
  "max": 10,
  "dimensions": {
    "accuracy": 9,
    "format": 8,
    "completeness": 9,
    "safety": 8
  },
  "notes": "Missing boundary condition for empty input"
}