Rubric-Based Prompts: How to Make AI Evaluate and Revise Drafts More Consistently

Asking AI to "improve writing" sounds simple, but the results are often inconsistent. Sometimes the language becomes neater, but too formal. At other times, AI removes important information or makes the writing sound like an advertisement.

The problem is not always with AI's capabilities. Often, we haven't provided a sufficiently clear measure of what constitutes a good result. This is where rubric-based prompts can help.

A rubric is a list of criteria used to assess a result. In the context of AI, a rubric can include aspects such as clarity, accuracy of information, writing style, structure, and completeness. AI is not only asked to generate or improve text but also to evaluate the text based on the standards we set.

Why do ordinary prompts often result in inconsistent revisions?

Prompts like “Please make this writing better” are too open-ended. The term “better” can mean shorter, more persuasive, more formal, or easier to read. AI has to guess our intent from limited context.

As a result, there are several common issues:

• The writing style strays too far from the original version.
• AI corrects grammar but overlooks facts or logical flow.
• The suggestions provided are too general and hard to apply.
• The revised results look polished but are not suitable for a specific audience or purpose.
• We have to repeat the same instructions in every conversation.

A rubric helps transform vague standards into criteria that can be checked one by one.

A simple rubric that can be used for various tasks

You don't need to create a complicated assessment system. For many needs, the following five criteria are sufficient:

1. Purpose: Does the writing achieve the requested goal?
2. Audience: Is the language appropriate for the intended readers?
3. Clarity: Is the main idea easy to understand?
4. Completeness: Is all important information available?
5. Style: Is the tone and format appropriate for the context?

Each criterion can be scored, for example, from 1 to 5. However, numbers alone are not enough. Add a brief explanation of what each score means so that AI does not interpret the scale arbitrarily.

For example:

• 5: Very clear, complete, and ready to use.
• 3: Good enough, but still has parts that need improvement.
• 1: Does not meet the goal or requires significant rewriting.

Rubric-based prompt template

Here is a template you can copy and customize:

You act as a meticulous editor.
Your tasks:
Score the draft based on the rubric below.
Briefly explain the reasons for each score.
Highlight the parts that need the most improvement.
Create a revised version without changing the main intent.
Assessment rubric:
Purpose: score 5 if the draft clearly addresses the main goal.
Audience: score 5 if the language is suitable for a general audience.
Clarity: score 5 if there are no confusing sentences.
Completeness: score 5 if all important information is available.
Style: score 5 if the tone is warm, concise, and not excessive.
Revision rules:
Maintain facts and the main intent.
Do not add facts that are not in the draft.
If there is missing information, mark it as [NEEDS COMPLETION].
Do not use technical terms without explanation.
Provide output in the following order:
Brief scoring table.
Main issues.
Revised version.
List of significant changes.
Draft:
[paste draft here]

This template can be used for articles, emails, product descriptions, proposals, and even internal documentation. The most important part is the relationship between the criteria, revision rules, and output format.

Example usage for work emails

For instance, if you have an email that is too long and want to make it clearer without sounding rude, the rubric can be modified to:

• Purpose: the recipient understands the requested action.
• Clarity: the main request appears at the beginning.
• Tone: polite, direct, and not passive-aggressive.
• Conciseness: no unnecessary repetition.
• Completeness: deadlines and important context are retained.

With these criteria, AI does not just shorten the email. It has clearer boundaries: do not remove deadlines, do not soften the request too much, and do not add reasons that we never wrote.

Use two stages: score first, revise later

One common mistake is asking AI to evaluate and immediately rewrite without a pause. The results may seem smooth, but we find it hard to understand the reasons for the changes made.

For important tasks, use two stages:

1. Evaluation stage: ask AI to provide scores, reasons, and a list of issues without rewriting.
2. Revision stage: after reading the evaluation results, ask AI to improve specific parts based on the agreed criteria.

This separation makes the process easier to audit. You can reject certain suggestions before a new version is created. This method is also useful when you want to maintain the original author's voice.

Do not treat scores as absolute truth

AI's scores are not objective measurements like calculator results. The scores are merely tools to identify weaknesses that may have been overlooked.

AI can still misjudge context, fail to capture humor, or consider certain sentences unclear when they were intentionally crafted for a specific audience. Therefore, do not blindly pursue a score of 5 in all categories. Good writing does not always have to be the most formal, the shortest, or the most complete.

Use the rubric as a checklist, not as a substitute for human judgment. For materials containing important data, business claims, legal information, or health advice, manual checks are still necessary.

What you can do now

Take one repetitive task that you often give to AI, such as editing emails or tidying up minutes. Then follow these three steps:

1. Write down the characteristics of the results you consider good.
2. Transform those characteristics into 3–5 criteria that can be checked.
3. Add rules about what should not be changed or added.

After several uses, pay attention to which criteria most often yield useful suggestions. The rubric does not have to be perfect from the start. In fact, the rubric will improve after being used on real examples.

In essence, a good prompt does not just tell AI what to create. A more mature prompt also explains how the results will be evaluated, what boundaries must be maintained, and in what format suggestions should be provided.

– Rio Yotto @rioyotto