Most people do not need complicated prompt formulas to get useful results from AI. They need a repeatable way to explain the task clearly, provide the right context and check the result before using it.
This tutorial shows you a practical prompting workflow you can reuse for writing, research, planning, marketing, analysis and everyday business tasks.
What you will learn
By the end of this tutorial, you will know how to:
- turn a vague request into a useful AI prompt;
- give the model the context it actually needs;
- define requirements without making the prompt unnecessarily complicated;
- specify the output format;
- use examples when precision matters;
- refine weak results with follow-up instructions; and
- review AI output before using it in real work.
Before you start
Choose one real task. Prompting becomes much easier when you work toward a concrete result rather than trying to create a universal "perfect prompt."
For this tutorial, we will use a simple example:
Write an article about Rovinj.
The request is understandable, but it leaves many important decisions to the AI. Who is the reader? What is the purpose? How long should it be? What should it cover? What format should it use?
We can improve it step by step.
Step 1: Define the task
Start by saying exactly what you want the AI to do.
A useful task statement normally contains an action and a deliverable.
Instead of:
Write about Rovinj.
Try:
Write a practical travel article about the best things to do in Rovinj, Croatia, for first-time visitors.
The second version establishes the deliverable, topic and audience immediately.
This principle applies to other tasks too:
- "Analyze these customer reviews and identify recurring complaints."
- "Rewrite this product description for a non-technical audience."
- "Create a launch checklist for a small WooCommerce store."
- "Compare these three hosting plans using price, storage, backups and support."
Do not add complexity simply to make a prompt look sophisticated. Clear instructions are more useful than decorative prompt language.
Step 2: Add relevant context
Context explains the circumstances around the task.
For our example, we might add:
The article is for an independent Croatia travel website. The audience is international visitors planning their first trip to Rovinj. They want practical information rather than promotional language.
Useful context can include:
- who the audience is;
- what the content will be used for;
- what the reader already knows;
- relevant business information;
- source material;
- constraints;
- definitions specific to your organization; and
- examples of previous work.
Only provide context that helps with the task. A large amount of unrelated information can make a prompt harder to maintain and review.
When factual accuracy matters
If the task depends on current prices, laws, product specifications, schedules or other changing information, tell the AI that current sources are required and verify important claims yourself.
For internal business tasks, provide the relevant source material instead of expecting the model to know private information.
Step 3: Define requirements
Now describe what a successful result should contain.
For the Rovinj example:
Include an introduction, 8-10 worthwhile things to do, practical tips, a one-day sample itinerary and a short FAQ. Aim for approximately 1,500 words. Use a friendly, knowledgeable and practical tone. Avoid exaggerated tourism language.
Requirements can cover:
- approximate length;
- topics that must be included;
- topics that must be excluded;
- tone;
- reading level;
- technical depth;
- geographic scope;
- dates or time periods;
- required evidence; and
- important constraints.
Avoid contradictory requirements
Prompts often fail because the instructions conflict.
For example:
Explain everything in detail, keep it under 300 words, include 15 examples and provide a comprehensive FAQ.
The model has no clean way to satisfy all four goals.
Decide what matters most and make the constraints realistic.
Step 4: Specify the output format
If you know how you want the answer organized, say so.
For an article, you might request:
Use a short introduction followed by descriptive H2 sections. Use bullet points only where they improve readability. Finish with an FAQ containing five concise questions and answers.
For other tasks, you could request:
- a table;
- a checklist;
- numbered steps;
- JSON;
- CSV-compatible rows;
- a short executive summary followed by detailed analysis;
- email copy;
- a comparison matrix; or
- a list of recommendations with supporting reasons.
Specifying structure is particularly useful when you plan to reuse the output in another workflow.
Step 5: Combine the pieces into one strong prompt
Our original prompt was:
Write an article about Rovinj.
The improved version becomes:
Write a practical travel article about the best things to do in Rovinj, Croatia, for first-time visitors.
The article is for an independent Croatia travel website. The audience is international visitors planning their first trip to Rovinj. They want useful local information rather than exaggerated promotional copy.
Include an introduction, 8-10 worthwhile things to do, practical planning tips, a one-day sample itinerary and a short FAQ. Aim for approximately 1,500 words. Use a friendly, knowledgeable and practical tone.
Organize the article with descriptive H2 and H3 headings. Use bullet points only when they make information easier to scan. Do not invent prices, opening hours or other details that require current verification. Clearly flag information that should be checked against a current source.
Notice that this prompt is not built from secret commands. It simply removes ambiguity.
Step 6: Add source material when the task needs it
For many business tasks, the best improvement is not another instruction. It is better input.
Suppose you want AI to rewrite a product page. Include the actual product specifications, customer type, brand guidelines and claims that are allowed.
A useful pattern is:
TASK
Rewrite the product description below.AUDIENCE
Small accommodation owners who are not technical.REQUIREMENTS
Keep all factual specifications unchanged. Explain benefits in plain English. Do not invent features.SOURCE MATERIAL
[paste verified product information]OUTPUT
Return a 100-150 word description followed by five concise feature bullets.
This separates instructions from the material the model should work with.
Step 7: Use examples when style or structure is difficult to describe
Sometimes showing the desired result is faster than explaining it.
For example:
Use headings in this style:
Weak: "SEO"
Preferred: "How to Check Whether Google Can Index Your Page"
Examples are particularly helpful for:
- classification;
- data extraction;
- naming conventions;
- formatting;
- brand voice;
- recurring reports; and
- structured content.
Do not overload a simple task with examples when the expected result is already obvious.
Step 8: Review the first result instead of automatically accepting it
A strong prompt can improve the first response, but prompting is normally iterative.
Review the output against the actual goal.
Ask:
- Did it complete the requested task?
- Is anything important missing?
- Did it misunderstand the audience?
- Are factual claims supported?
- Did it follow the requested structure?
- Is any section generic or repetitive?
- Did it invent information?
- Is the result usable without major cleanup?
Treat AI output as a draft that needs an appropriate level of review.
Step 9: Refine with targeted follow-up instructions
If most of the answer is good, do not restart from zero.
Tell the AI exactly what needs changing.
Examples:
Keep the structure, but make the practical tips more specific for someone visiting without a car.
Rewrite only the introduction. Remove promotional adjectives and get to the useful information faster.
The comparison is missing backup retention periods. Add that criterion without changing the existing columns.
Check your answer against the source material and list any claims that are not supported by it.
Targeted follow-ups preserve useful work while correcting specific weaknesses.
Step 10: Build a reusable prompt template
When a prompt works for a recurring task, turn its stable parts into a template.
Reusable business prompt template
Task: [What should the AI produce or do?]
Goal: [What outcome should this support?]
Audience: [Who will use or read the result?]
Context: [Relevant background]
Source material: [Verified information or documents]
Requirements: [Must include, must avoid, length, tone, constraints]
Output format: [Headings, table, JSON, checklist, etc.]
Quality check: [What should be verified before the answer is considered complete?]
You do not need every field for every prompt. Remove fields that do not contribute to the task.
A complete example: turning notes into a useful business article
Imagine you have these rough notes:
Backups important. Daily backup. Offsite. Test restore. Database + files. Before updates.
A weak prompt would be:
Turn this into a blog post.
A stronger prompt would be:
Create a beginner-friendly tutorial explaining how small WordPress website owners should think about backups.
Use these notes as required concepts:
– backups should include the database and website files;
– keep an independent or offsite copy where appropriate;
– backup frequency should reflect how often the site changes;
– create a backup before risky updates or migrations;
– a backup is only useful if it can be restored.Do not claim that one backup schedule is correct for every website. Explain the reasoning behind choosing frequency.
Structure the result as a short introduction, preparation section, numbered workflow, verification checklist and common mistakes section.
Write for a business owner who uses WordPress but does not administer servers.
The stronger version gives the model enough information to make useful decisions without dictating every sentence.
Common prompting mistakes
Asking for "the best" without criteria
"Give me the best hosting provider" leaves the evaluation criteria undefined.
Instead, define what matters: budget, location, traffic, support, control panel, backups or another requirement.
Using a role instead of real context
"Act as the world’s greatest SEO expert" is less useful than explaining the website, audience, goal, constraints and data available.
A role can help establish perspective, but it should not replace task information.
Asking for facts without providing or requesting sources
AI can produce fluent statements that still need verification.
For consequential or changing facts, require reliable sources and inspect them.
Adding dozens of unnecessary rules
More instructions do not automatically create a better result. Every rule creates another condition the response has to satisfy.
Use the smallest set of instructions that clearly defines success.
Mixing several unrelated tasks
If you ask for keyword research, a 3,000-word article, social posts, an email campaign and a technical SEO audit in one prompt, quality and controllability can suffer.
Break large workflows into logical stages.
Treating the first response as final
The conversation itself is useful. Point out what is missing, supply new information and ask for a targeted revision.
A practical quality-control prompt
After receiving an important draft, you can use a second instruction:
Review the response against my original requirements. Identify missing requirements, unsupported factual claims, vague sections and unnecessary repetition. Do not rewrite yet. Return a concise list of issues and the specific changes needed.
After reviewing that diagnosis, ask for the revision.
This separates evaluation from rewriting and makes it easier to see what changed.
Prompting for different types of work
Writing
Provide audience, purpose, source material, tone, required sections and claims that must not be invented.
Research
Define the question, date range, acceptable sources, geographic scope and how sources should be cited. Verify important findings.
Data analysis
Explain what each field means, define the question, state how missing values should be handled and specify the required output.
Coding
Describe the environment, expected behavior, constraints, relevant files and tests. Ask the model to verify the implementation rather than stopping after writing code.
Brainstorming
State the problem and constraints, but allow more freedom. Ask for genuinely different approaches rather than superficial variations.
A simple prompt formula to remember
When you do not know where to start, use:
Task + Context + Requirements + Output
For example:
Task: Create a homepage outline.
Context: This is for a small accounting firm serving Croatian SMEs. The main goal is qualified enquiries.
Requirements: Include trust elements, services, process, FAQs and one primary call to action. Avoid generic marketing slogans.
Output: Return the page sections in order. For each section, explain its purpose and suggest one headline.
That is enough structure for many everyday tasks.
Final checklist
Before sending an important prompt, check:
- Is the task explicit?
- Is the audience clear when it matters?
- Did I provide relevant context?
- Did I include necessary source material?
- Are the constraints realistic and non-contradictory?
- Did I specify the desired format?
- Have I distinguished facts from assumptions?
- Will I be able to judge whether the output is good?
- Does important information need independent verification?
Frequently asked questions
Do longer prompts always produce better answers?
No. A prompt should contain enough information to define the task clearly. Extra instructions that do not affect the result can create noise and make recurring prompts harder to maintain.
Should every prompt assign the AI a role?
No. A role can sometimes establish a useful perspective, but the task, context and requirements are normally more important.
Should I use one giant prompt or several smaller prompts?
Use one prompt when the task is coherent and the output can reasonably be evaluated as a whole. Split complex workflows into stages when each stage produces information needed by the next.
Can I reuse the same prompt?
Yes. Reusable templates are valuable for recurring work, but keep the variable information separate and review the template as your process changes.
How do I know whether a prompt is good?
Judge the output against explicit criteria. A good prompt is not the one that sounds most technical; it is one that reliably produces useful results for the task.
Where to go next
Once this basic workflow feels natural, improve prompts by adding better source material, examples and quality checks rather than simply making them longer.
For current platform-specific guidance, consult the official OpenAI prompt engineering documentation and help resources.
