Many people use one long prompt to ask AI to read materials, understand context, make decisions, write drafts, and check results all at once. This approach may seem practical, but it often produces answers that are hard to use: too general, missing important details, or only partially following instructions.
The solution is not always to create a longer prompt. Often, results will be better if the work is broken down into smaller stages. Imagine you are asking a coworker for help. You wouldn’t just say, “Read all these documents, find the issues, create a strategy, write a report, and then make sure everything is correct” without giving them a chance to check each step.
The following four-stage pattern can be used for light research, content creation, document analysis, idea generation, and even administrative tasks.
1. First Stage: Ask AI to Understand the Material
Don’t ask for the final result right away. Start by asking AI to read and summarize the provided material. The goal is not to get beautifully written text, but to ensure the main information is mapped out.
Example prompt:
“Read the following material. Identify the main topics, important facts, terms that need clarification, and any ambiguous parts. Do not make conclusions or recommendations yet. Present the results in four sections: summary, important facts, terms, and items that need clarification.”
This stage is useful because you can see if AI understands the context correctly. If the summary is already incorrect, that mistake can be corrected before moving on to the next stage.
For long documents, add constraints such as the number of points, date ranges, or sections of the document that should be prioritized. These constraints help reduce overly broad answers.
2. Second Stage: Turn Understanding into a Framework
Once the content is mapped out, ask AI to create a structure for the work. At this stage, don’t ask for the final paragraph yet. Request a list of arguments, a sequence of steps, or decision criteria.
For example, to create an article:
“Based on the summary, create three alternative article outlines. Each outline should include the main issue, core explanation, practical examples, risks or limitations, and steps readers can take. Mark information that comes directly from the material and sections that are proposals.”
This way, you don’t just receive one direction for writing. You can compare several structures before choosing the most suitable one.
The important thing is to separate facts from proposals. AI can help outline possible approaches, but those proposals still need human evaluation. A neat-looking framework may not necessarily align with the goals, audience, or actual conditions.
3. Third Stage: Request Results in a Clear Format
A common mistake in prompts is asking for “something nice” without explaining what the desired output should look like. Words like nice, professional, complete, and engaging have different meanings for everyone.
Replace abstract instructions with criteria that can be checked. For example:
- Use natural Indonesian that is not overly formal.
- Start with a problem that is close to the reader's experience.
- Use subheadings that are easy to scan.
- Provide one concrete example for each main concept.
- Explain technical terms when they first appear.
- Do not make claims that are not supported by the material.
- End with three actionable steps that can be taken today.
You can also explicitly specify the output format. For example, request a table with columns for “problem,” “cause,” “impact,” and “action.” For data that will be processed further, use a consistent format like JSON or a numbered list.
The clearer the expected output format, the less likely you will need to tidy up the answer from the start.
4. Fourth Stage: Conduct a Separate Review
Don’t ask AI to write while also stating that the writing is correct. Review is more useful when done as a separate task with clear criteria.
Example review prompt:
“Review the following draft using this checklist: 1) do all facts have a basis from the original material, 2) are there any claims that are too certain, 3) are there any repetitive sections, 4) are the reader's instructions clear, and 5) are there any terms that have not been explained? Create a table containing findings, locations, risk levels, and suggestions for improvement. Do not rewrite the entire draft.”
The instruction “do not rewrite the entire draft” is important if you want to receive an audit, not a new version that is difficult to compare with the original manuscript.
For more sensitive tasks, add human review. For example, if AI helps summarize contracts, financial reports, customer data, or health information, the results should not be considered final decisions just because they have passed an automated review.
Complete Workflow Example You Can Use Right Away
Here’s a brief pattern you can copy and adjust:
Stage 1 — Understand the material
Read the following material and map out facts, objectives, important terms, and unclear information. Do not make recommendations.
Stage 2 — Create a framework
Create two alternative structures based on the previous mapping. Separate facts from assumptions and proposals.
Stage 3 — Draft
Choose the best structure. Write a draft using clear language, concrete examples, informative subheadings, and practical conclusions.
Stage 4 — Audit
Review the draft based on accuracy, completeness, clarity, repetition, and overly certain claims. Present findings before making revisions.You don’t have to manually go through all four stages every time. For simple tasks, the first and second stages may be sufficient. However, the more important the results, the more beneficial it is to separate each stage.
What Does This Mean for Us?
Breaking down prompts is not just a trick to get longer answers. The main goal is to make the workflow easier to track. When the results are wrong, you can determine whether the issue arose during understanding the material, structuring, writing, or reviewing.
This pattern also helps reduce the tendency to accept the first answer as the final answer. AI can speed up work, but quality still depends on how we provide context, set criteria, and review results.
What You Can Do Now
- Choose one routine task that often results in messy answers.
- Break that task down into understanding the material, structuring, creating output, and reviewing.
- Write observable result criteria, not just words like “nice” or “professional.”
- Save successful prompts and note which parts still need improvement.
With this simple workflow, you won’t have to keep guessing magical sentences for prompts. What’s more important is to build a process that makes AI work in understandable and correctable steps.
– Rio Yotto @rioyotto