| name | eval-engineer |
| description | Define and maintain lightweight evaluations (acceptance criteria, regression checks, golden tests) so "green" is objective and repeatable. |
Eval Engineer — Make "Green" Objective
Use this skill when:
- A change risks regressions
- "Done" is ambiguous
- You need to justify that work is green with minimal cost
This skill is tool-agnostic. It produces verifiable checks and a cheap evaluation loop.
Principles
- Small and repeatable beats "big and rarely run".
- Prefer fast Tier 1 checks; reserve Tier 2/E2E for changes that demand it.
- Evals must be objective: either pass/fail or clearly measurable.
Outputs you must produce (minimum)
-
Acceptance criteria (2–5 items)
Each AC must be verifiable. Avoid "works" or "seems fine".
-
Regression risk statement (1 paragraph)
What could break, and what signal will catch it?
-
Evaluation loop
- Tier 1: fastest check(s) you will run every time
- Tier 2: when needed (integration/E2E or real environment)
AC templates (pick one; keep it short)
Bullet AC (default)
- AC1: ___ (how to verify)
- AC2: ___ (how to verify)
BDD-lite (when behavior is user-facing)
- Given ___
- When ___
- Then ___ (verifiable)
"Golden signals" checklist (cheap)
Pick the smallest set that fits:
- Integrity: static checks (lint/parse/schema)
- Behavior: one targeted scenario (manual or scripted)
- Regression: verify one previously failing bug stays fixed
Anti-patterns
- Writing evals that are too slow to run
- Treating logs as proof without a pass/fail signal
- Adding evals without documenting when to run them
