Many people try to improve AI results by adding more and more instructions in a single prompt. The problem is that overly long prompts often obscure the main objective. AI receives a lot of context at once, but it doesn't always know which parts are most important.
For tasks like writing articles, summarizing documents, creating plans, or analyzing data, a more practical approach is to create a workflow prompt: a series of small instructions executed step by step. It's like cooking. You wouldn't ask someone to "serve delicious food" without explaining the ingredients, the order of operations, and how to check the results.
Why does a single prompt often yield unstable answers?
A large prompt usually tries to do too many things at once. For example, a user might ask AI to read raw materials, determine the angle of the writing, create a structure, write the article, fact-check, refine the style, and prepare a title all in one command.
The result may look neat, but it's hard to identify which part is incorrect. Is the angle of the article off? Is the information incomplete? Or is the writing style not suitable for the audience?
By breaking down tasks, each stage has a clearer objective. You can also correct the direction before AI moves on to the next stage.
A four-stage pattern for everyday AI tasks
The following simple workflow can be applied to many types of work:
- Understanding the materials: ask AI to identify key information, objectives, and data gaps.
- Determining the plan: ask AI to create a structure or action steps before generating the final output.
- Working on the main part: use the approved plan as a foundation.
- Reviewing and refining: ask AI to audit the results against clear criteria.
These four stages do not always have to be conducted in separate conversations. However, separating them into several messages usually makes the process easier to control.
Step 1: Ask AI to understand the problem before answering
Don't immediately ask for the final result. Start by asking AI to read the context and demonstrate its understanding.
You will help me draft an article for a general audience. Before writing, identify: the main objective of the article, who the audience is, the three most important pieces of information, any unproven assumptions, and any missing data. Do not write the article yet.
Prompts like this are useful because they force the process to start with diagnosis. If AI misunderstands the objective, you can correct it early on.
For tasks involving documents, add clear constraints. For example, ask AI to distinguish between information written in the document and conclusions it draws on its own.
Use only the information from the text I provide. If there are conclusions or assumptions, mark them as inferences. If information is not available, write "data not available" and do not fabricate.
Step 2: Create a plan before generating output
Once AI understands the problem, ask for a work structure. This can be a list of subheadings, a decision table, a sequence of actions, or evaluation criteria.
Based on the previous analysis, create three alternative article structures. Each structure should have a target audience, main idea, order of subheadings, and potential misunderstandings. Do not write complete paragraphs.
The part "do not write complete paragraphs" is important. Without this constraint, AI often skips the planning stage and directly generates lengthy text.
If you already have preferences, ask AI to compare the options. For example:
Compare the three structures based on clarity, depth, and ease of implementation. Provide a recommendation for one option and explain the reasons in a maximum of five points.
Here, AI acts as a thinking partner, not just a writing machine.
Step 3: Work with concrete constraints
After the plan is chosen, only then ask AI to generate the main part. Avoid abstract instructions like "make it good" or "make it more professional." Replace them with verifiable criteria.
For example, instead of writing "create an engaging article," use:
- Start with a problem commonly faced by readers.
- Use natural and not overly formal Indonesian.
- Explain technical terms when they first appear.
- Include one example that can be tried immediately.
- Avoid unsupported claims.
You can also specify the output format to make it easier to transfer the results to other applications.
Write the results in the following format: title, one-paragraph summary, main steps, implementation example, and a checklist. Do not add sections outside of this format.
The clearer the desired output format, the less additional work there will be to tidy up the response.
Step 4: Make AI a reviewer, not just a creator
One common mistake is asking AI to evaluate its own results with simple questions like "is this good enough?" That question is too broad. Create a specific checklist.
Audit the following text against five criteria: accuracy to the brief, clarity for a general audience, presence of unsupported claims, repetition of ideas, and sentences that could be misunderstood. Create a table containing short quotes, issues, risk levels, and suggestions for improvement. Do not rewrite the entire text.
Requesting a report on issues first helps you understand what needs to be improved. After that, you can ask for targeted revisions:
Only fix the parts marked as high risk. Maintain the structure, tone, and correct information. After revisions, explain the changes made.
This approach is safer than asking AI to rewrite everything, as a total rewrite can alter parts that are actually correct.
Example workflow for creating an important email
For instance, you need to send an email to a client about project delays. The workflow can be structured as follows:
- Ask AI to summarize the facts: completed work, obstacles, impacts, and new dates.
- Ask AI to create three tone approaches: formal, warm, and direct.
- Choose one approach, then ask for a draft email with a subject, opening, explanation, solution, and call to action.
- Ask AI to check if the email sounds like it blames others or makes uncertain promises.
With this sequence, you not only get a neater email. You also reduce the risk of conveying unconfirmed information.
What does this mean for us?
Workflow prompts are not a way to make AI always correct. Their main benefit is to make the work process more visible. You can identify at which stage errors occur, what information is still lacking, and which decisions still require human judgment.
Use a simple rule: AI can assist in processing, comparing, and organizing information, but important decisions still need to be checked by someone who understands the context. Especially for legal, health, financial, security, or high-impact communication matters.
What you can do now
- Choose one repetitive task you often do with AI.
- Break that task down into understanding, planning, execution, and review.
- Write observable success criteria, not just "good" or "professional."
- Ask AI to mention assumptions and data that are not available.
- Save successful workflows as templates for future tasks.
Effective prompts are not always the longest prompts. Often, better results come from simpler instructions, but given in the right order.
– Rio Yotto @rioyotto