---
title: "One Prompt Can Shake Your Grades: Principles for Designing a Grading Prompt"
description: "The same answer can earn a different score depending on the prompt. Here are the principles for building a grading prompt that holds steady."
lang: "en"
date: "2026-04-12T15:00:00.000Z"
updated: "2026-06-28T02:37:00.000Z"
tags: ["grading prompts","ai assessment","automated grading","scoring reliability"]
canonical: "https://flipsson.com/en/blog/6NyhB2LglzoB"
---

# One Prompt Can Shake Your Grades: Principles for Designing a Grading Prompt

If you have ever graded the same student response twice and gotten two different scores, the problem is most likely not the response but ==an ambiguous prompt==. Three points the first time and two on a rerun means the criteria are shifting. A grading prompt is the constitution of your assessment. Write it vaguely and the results come out vague too. Before you switch to a pricier model, the right order is to sharpen the prompt first.

## A shaky prompt versus a steady one

"Grade this response" is the worst possible instruction. You are leaving the machine to assign a score without knowing what full marks even means. A steady prompt spells out the following.

- **What earns full marks**: Enumerate the elements that have to be present for a full score.
- **Partial credit rules**: Set arithmetically ==how many points each element is worth==.
- **What not to deduct for**: Make clear what is not being assessed, such as spelling, handwriting, or style of expression.
- **Output format**: Receive the score, the reasoning, and the location of the quoted evidence in a fixed structure.

==Making the partial credit rules arithmetic== is especially central to consistency. Phrasing like "partial credit if it is somewhat correct" gets read differently every time. Write it so it can be counted, as in "2 of the 3 core concepts present earns 2/3 credit."

## A procedure for checking stability

1. Grade the same 5 responses three times each.
2. If a score on the same response shifts, make the criterion wording more specific. Whatever shifts is precisely what is ambiguous.
3. ==Pin borderline cases, the genuinely ambiguous responses, right into the prompt as examples==.

> A good grading prompt is finished by a few examples of "a response like this gets this score." Examples are stronger than rules.

## Traps people fall into

- An overly long prompt gets its later instructions ignored. Put the core rules up front. The more important something is, the earlier you write it.
- Adjectives like "strictly" and "generously" are not criteria. Every person reads them differently. Convert them into numbers.
- Providing a model answer helps, but include a line like "credit equally valid answers too" so it does not harden into the only correct answer.
- Once you revise the prompt, rerun some previously graded responses and confirm the scores stay consistent.

## The skeleton of a good grading prompt

Do not try to write a perfect prompt from scratch; filling in the following skeleton is faster. Build the frame once and you can reuse it by swapping in a new item.

1. **Role and context**: Start by writing what subject, what grade, and what the item is assessing.
2. **Conditions for full marks**: List the elements that all have to be present for a full score.
3. **Partial credit table**: Lay out arithmetically how many points a given number of elements earns.
4. **Exclusions**: Write out plainly what you will not deduct for.
5. **Borderline examples**: Pin in two or three confusing responses along with the correct score for each.
6. **Output template**: Specify the format so you receive score, reasoning, and location of evidence in that order.

Fill in these six boxes and your scores hold steady in a way a vague one-line instruction cannot touch. Item 5 in particular drives the results, so keep growing the prompt by adding every ambiguous response you meet while grading as another example.

## Key takeaways

Consistency in grading comes not from model performance but from ==precision in the prompt==. Break full marks into elements, make partial credit arithmetic, and pin in borderline examples, and the scores stop shifting. A prompt you stabilize once gets reused as is next semester, so the time you spend refining it early becomes your longest-lasting asset.