Many people judge prompts by their length. However, a long prompt does not necessarily yield more useful answers. A more common issue is that we ask AI for something without explaining what constitutes a correct, complete, and usable result.
For example, the command "create a promotional caption" might produce an appealing sentence, but it could be too generic, not aligned with the brand's character, or forget to include a call to action. AI does not always fail to understand its task. Often, its success criteria have simply never been explained.
This is where the concept of pass criteria comes in handy. You not only tell AI what to create but also set a checklist that must be met before its answer is submitted.
This approach aligns with OpenAI's official guidelines, which suggest that users specify the task, context, goals, style, and output format clearly. Prompt engineering is not about finding a magic sentence, but about designing instructions that reduce the room for misinterpretation.
What are pass criteria?
Pass criteria are simple requirements that AI answers must meet. They can take the form of mandatory content, prohibitions, output formats, or ways to check the quality of the answers.
Consider asking someone to buy groceries. The command "buy ingredients for dinner" is still too vague. It would be more helpful to add: for four people, a maximum budget of Rp200,000, no seafood, and the result must be a shopping list categorized by item type.
In the prompt, these criteria turn a vague task into a job with verifiable objectives.
- Goal: what should be produced?
- Context: who will use it and for what purpose?
- Limitations: what should be avoided?
- Format: what should the final answer look like?
- Verification: how to ensure the answer meets the request?
Differentiate instructions, materials, and outcome standards
One practical way to structure prompts is to separate three parts: instructions, materials, and outcome standards. This separation helps AI distinguish what needs to be done from information that only needs to be analyzed.
1. Instructions
Clearly explain the role and task of AI. Avoid overly long openings if they do not add context.
You are assisting the customer service team. Rewrite the following response to be clearer, more polite, and less defensive.2. Materials
Place the text, data, or notes that need to be processed in a separate section. Separators like triple quotes or clear labels can reduce confusion, especially if the materials contain other instructions.
CUSTOMER MESSAGE MATERIAL:
"""
The item that arrived is the wrong color. I have been chatting since yesterday but have not received a response.
"""3. Outcome standards
This is the part that is most often forgotten. Specify verifiable requirements, not just adjectives like “good” or “professional.”
Outcome standards:
- Maximum 80 words.
- Acknowledge the customer's issue in the first sentence.
- Do not blame the customer or the courier.
- Explain the next steps.
- End with a question that aids resolution.OpenAI also recommends separating instructions from context and specifying the output format clearly. This principle can be applied in both casual conversations and when building API-based applications. Their official guide provides examples of using labels and separators to make prompts easier for the model to understand.
Use verifiable criteria
Good criteria do not have to be complicated. The important thing is that the criteria can be answered with “yes” or “no.”
Compare the following two instructions:
- “Write something interesting and informative.”
- “Write 500–700 words, use three subheadings, include one concrete example, and do not make claims without sources.”
The second instruction is easier to evaluate. You can count the length, check the number of subheadings, verify the example, and flag claims that lack sources.
Some types of criteria that can be used include:
- Content criteria: three benefits, two risks, one example.
- Length criteria: maximum 150 words or five points.
- Style criteria: natural Indonesian, without unexplained jargon.
- Format criteria: table, numbered list, JSON, or paragraph.
- Safety criteria: mark uncertain parts and do not fabricate data.
Add verification steps before the final answer
For sufficiently important tasks, ask AI to perform an internal check before providing the final result. You do not have to request its entire thought process. Just ask for a summary of the checks or a brief validation list.
Before providing the final answer:
1. Check if all the above criteria are met.
2. Mark any information that is unavailable or still uncertain.
3. If there are unmet criteria, revise the answer.
4. Display only the final result and a brief note if necessary.This technique does not guarantee that the answer will always be correct. However, it makes the task more structured and gives the model a chance to identify obvious shortcomings, such as missing points or incorrect formats.
Prompt template you can try immediately
Here is a general template that you can customize for writing, analyzing documents, summarizing, or preparing office work:
Goal:
[explain the desired outcome]
Context:
[explain the reader, user, situation, or background]
Materials:
"""
[insert text, data, or notes here]
"""
Instructions:
- [main task]
- [additional tasks if any]
Pass criteria:
- [first mandatory content]
- [length or format limitation]
- [language style]
- [things to avoid]
- [how to handle unknown information]
Before answering, check all criteria. If any information is lacking, state it clearly and do not fabricate.What does this mean for us?
Better prompts do not always mean longer ones. What is more important is transferring the standards that exist in our minds into instructions that AI can read.
Start with one frequently repeated task, such as replying to emails, summarizing meetings, brainstorming content ideas, or turning notes into task lists. Write down three to five of the most important pass criteria. After receiving an answer, see which parts still miss the mark, and add new criteria if necessary.
In this way, you do not have to keep searching for new prompts. You are building a workflow pattern that can be improved over time—more like a quality control checklist than just a one-time command.
Sources & further reading
- Prompt engineering best practices for ChatGPT
- Best practices for prompt engineering with the OpenAI API
– Rio Yotto @rioyotto