eval-rubric-format-structure
Installation
SKILL.md
Eval Rubric — Format and Structure
When to use this
Apply when the format and presentation of a legal AI output needs to be assessed independently of its legal accuracy. A well-structured contract draft makes review faster; a poorly structured one wastes a lawyer's time. Format quality is a secondary rubric — it rarely changes whether a deployment is blocked, but it tracks the product's output polish over time.
Run in the [[eval-llm-as-judge-system-prompt]] ensemble as an optional rubric (weight it lower than legal soundness and citation quality).
Scoring (0–5)
| Score | Label | Criteria |
|---|---|---|
| 5 | Excellent | Output type perfectly matched to the request; correct heading hierarchy (numbered clauses for contracts, ##/### for analysis); tables used where appropriate (comparison, calculations); lists used for enumerable items; no unnecessary boilerplate; RTL/LTR correct for language; professional legal register throughout |
| 4 | Good | Well-structured with minor issues (e.g., inconsistent heading numbering, a table that could be a list) |
| 3 | Acceptable | Reasonably structured but with clear formatting gaps (heading levels mixed up; a comparison given as prose instead of a table; a contract clause not numbered) |
| 2 | Poor | Significant structural problems (no headings at all in a 2,000-word draft; a calculation result buried in prose; comparison jumbled) |
| 1 | Very poor | Nearly unstructured; would require complete reformatting to be usable |
| 0 | Fail | Completely inappropriate format for the request type (e.g., a JSON object returned for a contract draft request) |