I overheard some young employees at my company saying, “Prompt engineering is really important, isn’t it?”
I didn’t understand what they meant, so I couldn’t join the conversation.
I searched for it that night, but when I saw Python, machine learning, schools, and certifications lined up, I just closed the tab.
On the other hand, I use ChatGPT two or three times a month to draft proposals.
But the text that comes back is always vague, and I end up rewriting the whole thing myself.
This article starts by splitting the term “prompt engineering” into two. Once you decide which side you are on, the things you need to do to get started will decrease significantly.
What you can take away from this are four things: 1) Two questions to decide which side you are on, 2) the reason why 0-cost materials are enough if you are on the user side, 3) the 5 components you need to memorize, and 4) how to clear the three hurdles where self-study usually stalls.
To clarify my position upfront, this article does not recommend certifications or schools.
However, I will leave the conditions for who this is suitable for at the end.
The reason you get lost in self-study is that the same term refers to two different things
The reason you can’t find a clear answer when searching for how to study on your own is not because you are bad at searching.
It is because the same term refers to two different things.
At first, I also thought it was a problem with my own comprehension.
I would read articles for developers and back out thinking, “What is this about?” and then read articles on tips for asking questions and get lost wondering, “Is this the important skill?”
I did that back-and-forth about twice and closed the tabs without deciding which one I should be reading.
It wasn’t so much that I didn’t understand the meaning, but rather that I was mixing up texts intended for different audiences.
The technology of the creator and the etiquette of the user. They are different things
The first is the technology for the side that builds systems incorporating AI. What information to pass to the AI, what to have it reference, and how to verify the output returned.
This is a continuation of programming, and the reason Python results appear in searches is that there is a lot of information for this side.
The second is the etiquette for the side that asks AI for help with work. The question of what and how much to write to get the intended response.
I think it is closer to how to write instructions for an external contractor than it is to technology.
It is not a matter of which is superior.
It is just that the content you learn hardly overlaps. That is why there are few articles that explain both in one piece, and the reader gets lost.
The reason some articles say “Prompts are not the essence”
When you search, you will also find articles to the effect that “polishing prompts is meaningless.”
I read this and stopped for a moment.
That is because I thought there was no point in learning it.
But that is talking about the first side.
To make it work stably as a system, just knowing how to ask is not enough.
The context is that it won’t be useful unless you design how to place the information to be referenced and how to check the output.
It is true that such points of contention exist.
However, the intended audience is just different, and it is not something that a user who just wants to write their own documents needs to receive.
Which side you are on is decided by two questions
There are only two questions to decide.
-
Do you have plans to build AI tools or systems yourself within the next six months?
-
Are you asking the AI for documents for your own work, or features for others to use?
If the answer to the first is “no” and the second is “documents for my own work,” then you are on the second side.
If you are just an employee who wants to use AI for work, most people fall into this category.
The rest of this article will deal only with the second side.
For those aiming for the first, I will draw a line in the final chapter.
Decide which side you are on by looking at your schedule for the next six months.
For those on the user side, a $0 official guide is enough to get started
This is about how to choose learning materials.
To put it bluntly, books, schools, and certifications are things you should open later.
Even though the things you should open first are free, they don’t appear at the top of search results.
This is where you are most likely to waste money when teaching yourself.
Before buying materials, look at the provider of the AI you are using
The providers of each AI publish guides on how to prompt for free on their official websites.
For example, Google publishes their approach to building prompts using four elements—persona, task, background information, and format—in How to write AI prompts that are effective in business.
Microsoft also explains in Get started with prompt engineering in Microsoft Copilot how to build prompts using four elements: goal, context, source, and expected output.
If you try it, you’ll find that almost everything you need for an introduction is covered on these types of pages.
Since the official source is the first to update when specifications change, you are less likely to encounter outdated explanations.
If you are using ChatGPT, it is faster to open the official Help Center and search for “prompt.”
You can find them by searching for “service name + official help.”
Bookmarking prompt collections won’t help you learn to write them
I also bookmarked articles summarizing prompt collections.
As a result, I never used them once.
I found out why later.
The examples collected were written under the conditions of someone else’s project, so you need to rewrite them to apply them to your own project.
And the reason you can’t rewrite them is that you can’t distinguish which parts are conditions and which parts are fixed text.
It’s not that the collections themselves are bad.
They work well once you understand the components.
That’s why in the next chapter, I will break down prompts into components.
There are only 5 components to learn for beginners
This is the main part of the article.
I will narrow it down to a level you can start in one hour on a Saturday morning.
I will use a draft of a proposal that I was actually struggling with for all examples.
If I use different subjects for each component, you won’t be able to apply them to your own work.
Goal, Audience, Conditions, Format, Examples. Don’t worry about the order
These are the 5 things to remember.
-
Goal: What is the document used for? (e.g., A proposal for a new product to a construction firm)
-
Audience: Who is reading this? (e.g., a setting where both the president and the on-site craftsmen are present)
-
Conditions: Essential facts, figures, and constraints (e.g., construction time, price difference from existing products)
-
Format: Length, structure, bullet points vs. prose (e.g., one A4 page, three headings)
-
Example: Paste one document of the same type that has been approved in the past
The last one, the example, is the most effective.
Yet, it is the one most often omitted.
You don’t need a specific order or set phrases.
As long as you include these 5, you can write in your own words.
In other words, it is not something to be memorized.
Vague answers are usually the result of missing components 3 and 5.
AI fills in parts where conditions are not specified with safe, generalities.
So, if you get a generic response, it’s not a matter of writing talent, but a “sign that you haven’t provided the conditions.”
This is the kind of request I used to send for a long time.
この内容で、工務店向けの提案書を作ってください。
When you add the 5 components, it looks like this.
工務店の社長と職人が同席する場で読む提案書のたたき台を作ってください。
外せない条件は、施工時間が既存品より短いこと、価格は既存品より少し高いこと。
A4で1枚、見出しは3つまで、各見出しの下は箇条書きで。
文章の調子は、下に貼る去年の提案書に合わせてください。
（去年、同じ商品群で通った提案書をここに貼る）
There is no need to memorize long examples.
I only added two things: 3. Conditions and 5. Examples.
Since I started pasting documents that were approved in the past, the amount of rewriting I have to do has visibly decreased.
What surprised me was that what worked wasn’t clever phrasing, but documents I already had on hand.
Don’t try to get it right on the first try. How you send the second prompt is the real core.
This was the most misunderstood part of the introduction.
At first, I thought it was “practice in thinking of a clever sentence.”
In reality, it’s different.
Looking at what comes back and adding the missing conditions is the real core.
The pattern for how to correct it is simple.
-
Specify only one point that is off (e.g., “The explanation of the price is weak”)
-
Provide specific instructions on how to fix it (e.g., “Include the difference from the existing product and the payback period”)
-
Don’t make it redo everything; have it rewrite only that part
If it doesn’t get closer even after 3 tries, separate the issue by assuming the materials provided are insufficient, not the way you asked.
If you keep tweaking the wording there, it usually won’t come back right.
If you want to see the 5 components assembled into the format of your own work, I have compiled task-specific templates in the paid article “The Complete Guide to ChatGPT Prompt Tips | 20 Task-Specific Templates and How to Resolve Failures.”
It is perfectly fine to try these 5 components for free first.
Self-study stalls at three points before you even choose your materials
There are generally three places where you will get stuck.
1.Not knowing if it is okay to paste company information 2.Wasting time looking for practice topics 3.Not being able to judge for yourself if your output has improved.
I will cover 1 and 2 in this chapter.
Since 3 requires another person’s perspective, I will touch on it in the final Q&A.
Who to ask and what to ask before pasting client names and quote amounts
In sales or purchasing jobs, this is where you will inevitably get stuck.
I got stuck here too.
The bare minimum is not to include client names, personal information, or undisclosed prices and quantities.
Since rules vary by company, I cannot give a definitive answer here.
With that in mind, it is faster to decide on one question to ask your IT department.
Is there a list of AI tools that are approved for entering work-related information?
If your company does not have a list, make arrangements to get approval from your supervisor before using them.
I want to avoid the state of being unable to do anything until you ask, so here is an escape route.
If you replace the company name with “Company A” and the amount with “XX yen,” you can start practicing today.
You do not need real names just to see if the structure of the returned text changes.
Searching for practice topics will waste an hour of your Saturday
This was entirely my own failure.
The time spent thinking about “what should I practice” was the biggest waste.
There is only one way to decide:
choose one document from last week’s work that took a long time to rewrite.
Work on that same topic for 4 weeks.
There is a reason why you shouldn’t change the topic every time.
If you change it, you won’t know if you have improved compared to before.
With the same document, you can see the direct effect of adding conditions.
Also, do not make finishing a book your goal.
You will get a sense of progress faster by redoing one document five times.
Thinking about certifications, schools, and careers comes after this
Up to this point, we have been on the second side, which is the user’s perspective.
Finally, I will touch on the first side and the topic of certifications.
I do not recommend them.
However, I will leave the criteria for people who are suited for it, so you can confirm that it is not for you and move on.
Before deciding whether to make “Prompt Engineer” your profession
Job openings for this role are leaning toward the side that builds systems incorporating AI.
To be honest, it is not a job that can be done solely by being good at asking questions.
If you are aiming for this, you need to learn the first side.
Programming and how to refine the verification of outputs.
This is not an extension of spending one hour on a Saturday; it is a different kind of accumulation.
Conversely, if you are on the user side, the professional aspect is irrelevant.
I don’t think you need to panic by looking at numbers regarding future prospects or demand.
Here is when you might consider certifications and courses.
There are only two conditions.
-
You plan to be in a position to explain or promote it within the company (being able to use it yourself is not enough)
-
You need objective proof for a transfer or job change
If neither applies, your priority should be to cycle through the 5 components for 4 weeks.
Costs and schedules are subject to change, so I will not write down prices here.
Please check the official page of the organizing body after you have decided to take it.
I will not name specific schools either.
I don’t believe self-study is always the right answer, but this is a matter of order.
Summary: There are 5 components to learn, and the core is how you throw the second prompt.
Let’s review in 4 lines.
-
The reason you were lost is that the same term referred to both “the technology of the builder” and “the etiquette of the user”
-
If you are a user, the free official guide is enough for learning materials
-
There are 5 components to learn. The most effective one is examples
-
The core is not getting it right on the first try, but how you throw the second one
There is only one thing to do today:
choose one document that took the most time to rewrite last week, paste one document of the same type that was accepted in the past, and then ask again.
No registration or materials are required, so it will be finished in 10 minutes.
The next thing to decide is how far to expand your AI studies.
How to narrow down the scope and how to proceed over 4 weeks is summarized in Where should working adults start with AI study? Only 3 things to keep for self-study.
FAQ: About self-studying prompt engineering
Q1: How long does it take to get the hang of self-studying prompt engineering?
It is more accurate to think in terms of number of attempts rather than duration.
If you redo the same document 5 times, you will start to notice for yourself, “I missed a condition here.”
If you spend one hour on Saturdays, that is about one month’s worth.
I won’t give a number for how many hours it takes to master, as that varies too much from person to person.
You can verify the number of attempts yourself.
Q2: Can non-liberal arts majors or non-engineers do it?
If you are on the user side, yes.
Of the two types I distinguished in the main text, if you aim to be on the creator side, you will need programming, but if you are just using AI for your own work documents, you don’t need to go that far.
The reason Python and machine learning appear when you search is that there is more information available for the creator side.
It doesn’t mean it’s impossible for you, so please don’t stop there.
Q3: Should I use prompt collections or template articles?
They are effective once you understand the components.
If you collect them first, you will just end up with more bookmarks without being able to rewrite them for your own cases.
When using them, look at where the conditions are in the example.
If you replace those parts with your own numbers and constraints before sending, it will work properly.
It’s a matter of order; the example collections themselves are not bad.
Q4: Should I buy books? What kind of books should I choose?
First, start with the 0-yen official guide.
If you still want to buy one, there is only one way to tell them apart.
Look at the table of contents and judge whether it is for the creator side or the user side.
If there are many chapters on development, implementation, and APIs, it is for the creator side, so even if a user reads it, they won’t be able to put it into practice.
I will not list any book titles.
The core of this article is not to increase the number of study materials.
Q5: Does writing in English improve accuracy?
If you are creating Japanese business documents, writing in Japanese allows you to convey the conditions more accurately.
The phrasing of company names, industry terminology, and the sense of distance with clients.
The conditions will drop by the amount of effort it takes to translate these into English.
I will not assert that English is superior.
Since the extent of the difference requires verification, I will refrain from providing numbers.
Q6: Are there people for whom self-study is not enough?
Yes, there are.
People who cannot judge for themselves whether their output has improved need someone to look at it for them.
However, that person does not have to be from a course.
It is better to first have someone in your company who writes the same type of documents check if they get the same results.
I think it is fine to consider a course only if that is not enough.
Read also
I write one or two articles a week about money and working styles for office employees.
Please follow me if you like.
*This article was created using generative AI.
