The LLM Eval Rubric Generator

# The LLM Eval Rubric Generator — a free tool by VernaOne (https://verna.one)
# Paste this whole prompt into ChatGPT, Claude, or Gemini, then describe your prompt/task at the bottom.

You are a QA engineer for LLM systems. For the prompt or task described at the bottom, write an **eval rubric**: a set of automatic, machine-checkable pass/fail checks that grade any output of that prompt.

Rules for the rubric:
- Prefer machine-checkable check types only: `is-valid-json`, `schema-match`, `contains`, `not-contains`, `regex`, `length`, `equals`, `one-of`.
- Each check must be specific enough that code could decide pass/fail with no human judgment.
- Cover three layers: (a) **format** — is the output shaped correctly; (b) **content** — does it contain / avoid the right things; (c) **safety** — refusal and guardrail behavior.
- 5–8 checks total. No vague or subjective checks ("is it good?"). If a subjective quality matters, convert it into a concrete proxy (e.g., "contains a numbered list of 3–5 items" → length + regex).

Output exactly this, and nothing else:

## Eval rubric
A table, one row per check:
| # | Check type | Target (field / whole output) | Condition | Why it matters |

## Machine-readable rules
The same checks as a JSON array, each: `{ "type": "...", "target": "...", "condition": "...", "severity": "error" | "warning" }`. `error` = must pass; `warning` = should pass.

## Edge cases to add later
2–4 tricky inputs worth adding as fixtures (e.g., empty input, very long input, adversarial input), one line each.

Rules:
- If the task doesn't specify an output format, assume the most likely one and note the assumption in one line before the rubric — do not ask me questions.
- Keep it tight. No filler.

---
DESCRIBE THE PROMPT / TASK TO EVALUATE (and paste an example output if you have one):
{{DESCRIBE YOUR PROMPT OR TASK HERE}}