Many people improve AI results by writing increasingly lengthy prompts. However, the issue often isn't a lack of instructions, but rather the absence of clear standards for evaluating responses.
For example, if you ask AI to summarize a meeting, simply writing “Summarize these notes” may result in an answer that is too long, misses important decisions, or mixes facts with assumptions. AI needs a picture of what a good answer looks like.
This is where assessment rubrics come in handy. A rubric is a list of criteria used to measure the quality of results. In the context of AI, a rubric helps the model understand priorities, limitations, and the expected format of the output.
Why are ordinary instructions often not enough?
Instructions like “write something engaging” or “explain simply” sound clear to humans, but are still open to many interpretations. Engaging for whom? Simple to what extent? Is it permissible to use technical terms? How long should the writing be?
Without more concrete criteria, AI tends to fill in the gaps with general patterns from its training data. The results may look neat, but they may not fit your needs.
A rubric transforms subjective requests into several indicators that are easier to follow. This doesn't mean AI will always adhere to everything perfectly, but the chances of getting suitable results increase.
A simple rubric structure for prompts
You don't need to create a complicated assessment system. For many everyday tasks, the following four parts are sufficient:
- Objective: what should the answer achieve?
- Main criteria: what aspects are most important?
- Limitations: what should not be done?
- Final check: what should be checked before the answer is given?
For example, to create a summary of an article, the rubric could state: maintain the main facts, separate facts from opinions, use easily understandable Indonesian, a maximum of five paragraphs, and do not add information not present in the source.
Example prompt that can be used directly
Here is a template that you can customize:
Act as an editor who helps me compose an answer based on the provided material.
Objective: create a summary that helps readers understand the core material quickly.
Use the following criteria:
- Prioritize facts, decisions, and actionable steps.
- Use clear and concise language.
- Separate facts from opinions or interpretations.
- Do not add information not found in the material.
- Use subheadings if there are more than two topics.
Before providing the final answer, check if all the above criteria have been met. If there is missing information, clearly mark it instead of guessing.
Material:
[paste material here]
Note that this prompt not only explains the task. It also provides quality measures and requests a check before the answer is submitted.
Use priorities, not a long unordered list
The more criteria you add, the greater the likelihood of conflicts. You might ask for an answer that is brief, complete, casual, formal, technical, and easy to understand all at once. Not all of these goals can always be achieved at the same level.
Therefore, establish priorities. For example:
- Accuracy and fidelity to the source.
- Clarity of structure.
- Completeness of important information.
- Language style.
- Conciseness.
This order helps AI make decisions when it has to choose. If an answer cannot be both very complete and very brief at the same time, accuracy and important information should take precedence.
Add an output format that is easy to check
A rubric will be more useful if the final results have a consistent format. For example, to evaluate content ideas, you could request the following structure:
- Main idea
- Problem solved
- Target audience
- Risks or weaknesses
- Improvement suggestions
- Relevance score: 1–5
Such a format makes it easier for you to compare several results. AI not only generates long text but also presents information in a form that can be reviewed one by one.
However, scores from AI should be considered as a tool, not a final decision. The model can provide scores that seem objective, but the assessment still depends on the instructions and context you provide.
Differentiate assessment from facts
One risk of using rubrics is that AI may appear overly confident when assessing something that is actually subjective. For example, a score of “writing quality: 9 out of 10” does not automatically mean that the writing is correct, effective, or suitable for all readers.
To mitigate this issue, ask AI to explain the reasoning behind the assessment with concrete evidence. The prompt can be added as follows:
For each criterion, provide a status of met, partially met, or not met. Include a brief reason referring to parts of the answer. Do not give a score without explanation.
This approach makes the evaluation easier to audit. You can see whether the assessment is indeed supported by the content of the answer, rather than just a number that sounds convincing.
What you can do now
Choose one task that you often assign to AI, such as composing emails, organizing notes, drafting article ideas, or reviewing code. Then take the following three steps:
- Write the final result that you actually want.
- Note three to five common mistakes that frequently occur.
- Turn those mistakes into checking criteria in the prompt.
If AI often makes emails too long, add a limit on the number of paragraphs. If it frequently misses details, ask for a list of key points before drafting the final version. If it often guesses, instruct it to mark information that is not available.
After using it several times, the rubric can become a working template. You won’t have to start from scratch every time you open a new conversation.
Conclusion
A good prompt is not just a command, but also explains how a good result should be recognized. Assessment rubrics help AI understand priorities, reduce overly free interpretations, and make results more consistent.
However, rubrics are not a substitute for human review. For important decisions, especially those involving finance, law, health, security, or personal data, AI results still need to be verified with appropriate sources and considerations.
– Rio Yotto @rioyotto