If AI can't do what you want, it's probably because the requirements aren't clearly stated.

This article was originally published in Vista E-newsletter No. 148, Read the original e-newsletter (opens in a new tab).

Hello, hello!

In the last issue, I talked to you about my experience in the revision of vista.tw. After I sent it out, I received a lot of replies. There is one type of question that I have been asked dozens of times this year, with several different ways of saying the same thing:

“I also tried to ask AI to help me make something, but what it made was completely different from what I thought. After a few changes, it became more and more messy, and finally I had to give up.”

In this e-newsletter, I want to seriously answer this question. Well, let me tell you my own observations first: most of the time, the problem is not that the AI ​​is not smart enough, but that the sentence we give it leaves too many blanks for it to guess. There are ways to improve this, and it has nothing to do with whether you know how to program.

In the second half of the letter, I will also introduce you to a new attempt: the “Build your first AI gadget in seven days” challenge camp (opens in a new tab) that started at the end of November.

You can do something with just one sentence, but a lot of it is done by AI.

Well, let’s start with a very common opening.

Suppose you want to make a small tool that records the amount of water you drink every day, so you open a tool such as Claude or ChatGPT and type this sentence:

“Make me a water drinking record tool.”

I bet the AI will make something, and it usually looks pretty good. The problem is, there are at least seven or eight things hidden in this sentence that it must decide for itself:

• How many milliliters is a cup? Do you want to set it yourself?

-How many drinks a day is your goal?

• Use the button to add a cup, or enter a number?

• Should there be charts, history or reminders on the screen?

• Do you want to automatically reset to zero the next day?

• Is it just for you or for the whole family?

Of course, you actually already have the answer in your heart, but you just haven’t said it out loud. AI can’t read your mind, so it has to make decisions for you based on what an average person might want. The result is: it works hard, but you feel like nothing is right.

Next, most people will start to make corrections bit by bit: “The buttons are bigger”, “No charts”, “What I want is to start over every day”. Every time it is repaired, the AI ​​may change other places. By the fifth time, the originally good function is broken, and you have forgotten what you wanted first?

When I first started learning Vibe Coding (opens in a new tab), I was stuck in this cycle for a long time. Later, I gradually discovered that instead of fixing it all the time afterwards, it is better to spend five minutes writing down what was in my mind before speaking.

Five-paragraph explanation: Give the picture in your mind to AI

In the past few months, I have sorted out a very simple format. Before asking AI to make a gadget, I almost always write it down. There are five paragraphs in total, and two or three sentences in each paragraph are enough.

Paragraph 1: Who uses it and when

Example: “I use it myself. I open it once every morning when I get up and before going to bed.”

This section determines the basic tonality of the tool. If you are using it for yourself, it doesn’t matter if the interface is a bit cool; if you want to show it to customers, you have to pay attention to the layout and explanatory text. If you open it several times a day, the operation must be very fast; if you only use it once a month, a few more steps don’t matter.

Paragraph 2: Things to do

List three to five features, putting the most important first.

Example: “1. Click to add a drink. 2. The screen will display how many drinks you drank today. 3. It will automatically reset to zero at midnight every day. 4. You can view the records of the past seven days.”

The order is important. AI will put the most effort on the previous projects. You can also use this order to judge which functions are not acceptable and which ones can be added later.

Paragraph 3: What does the picture look like

Just describe it in your own words, no design jargon required.

Example: “The top is today’s cup count, with big fonts; the middle is a big round button; the bottom is the seven-day record, just use a simple long bar.”

If you have a favorite app or webpage at hand, you can also just say “the overall feeling is as clean as a certain app”, and the AI will usually get the direction.

Paragraph 4: Unwanted things

This paragraph is most often ignored, but it is the most useful.

Example: “No need to log in, no reminder notifications, no sharing functions, no animations.”

AI likes to add functions for you, which is a sign that it wants to do things well, but for a gadget, one more function means one more thing that may break. By telling what you don’t want first, you can save half of the back-and-forth revisions.

Paragraph 5: How to do it well

Give a standard to check against.

Example: “I can complete a recording in three seconds, and it won’t run when I open it on my mobile phone.”

With this section, after the AI is finished, you don’t have to judge whether it is good or not based on your feelings, you can just check it according to this standard.

After writing these five paragraphs, the entire specification is only about two to three hundred words. Paste the entire paragraph to AI, and add the sentence “Please follow this specification, and when you are done, tell me one by one where each function is done.” You will find that the first version is much closer to what you thought.

Complete demonstration: a quotation calculator for accepting cases

Just looking at the format may be a bit abstract, so let’s go through it with a complete example.

Suppose you are a graphic designer who takes a case. Every time a client asks, “How much does it cost to make a set of business cards and flyers?”, you have to open a spreadsheet, add items one by one, and then cut out a paragraph of text and paste it to the other party. You want to make a small tool that can do this in a minute.

One sentence version

“Help me build a design quotation calculator.”

AI quickly made it: there are more than a dozen service items on the screen, each item has a unit price column, and there is also a beautiful pie chart showing the proportion of each item in the total price. It looks very professional, but you realize something is wrong after two minutes:

• The project is what it imagined, not the case you are actually taking on

• You have to enter the unit price yourself every time, which is as troublesome as a spreadsheet

• There is no price increase for urgent items, but this is the most common situation you encounter.

• The pie chart looks great, but customers can’t see it at all.

• After finishing the calculation, you still have to write a paragraph and post it to the customer.

Next, you start to revise. By the fourth revision, the pie chart is removed and the price increase for urgent items is added, but the original subtotal begins to be miscalculated.

Version rewritten in five paragraphs

Take five minutes to write down the same requirement in five paragraphs:

• Who uses it and when to use it: I use it myself. When customers inquire about prices on LINE, they open it on their mobile phones and must be able to reply within one minute.

• Things to do: 1. Check the service items (business card design 3,000, flyer design 4,500, poster design 6,000, Logo design 12,000), and write the price in the program first. 2. You can check “urgent” and add 30% to the total price. 3. Automatically calculate the total price before tax and the total price including tax. 4. Generate a quotation text that can be copied directly.

• What the screen looks like: one item per line, a check box on the left, and a price on the right; Urgent is the switch at the bottom; the total price is in big letters at the bottom; and there is a “Copy Quotation Text” button at the bottom.

• Things you don’t want: no charts, no login, no customer database, no mail-sending function.

• How to calculate it well: After the customer asks for the price, I can check, copy, and paste it into LINE within one minute; the format of the quotation text is “Hello, business card design and flyer design, total NT$7,500 (excluding tax), tax included NT$7,875, it will be delivered in about five working days.”

Paste this paragraph to AI, and the first version will look like what you want: four items, a rush switch, a copy button, and nothing superfluous. You only need to fine-tune one or two things, such as changing the delivery days to optional.

What is the difference between the two versions

If you compare it carefully, you will find that the five-paragraph version does not use any technical terms. You just write down what is already in your mind: the real service items and prices, the most common emergency situations, the scenarios you actually use, and what you want the final piece of text to look like.

The most critical of these are two things. First, the price and quotation text examples are all your own real data, and the AI ​​does not need to edit them. Second, the “unwanted things” section directly blocks functions such as pie charts that look powerful but you don’t use at all.

This is why I often say that when asking AI to make things, the most important ability is not technology, but your understanding of your job. The clearer you are about how you work every day, the clearer you can articulate your needs.

The three most common stuck levels and how I deal with them

No matter how clearly the specifications are written, there will still be obstacles in the process. The following are the three situations that I encounter most often and are most commonly asked by students.

Situation 1: A was changed, but B was broken

This is the most common one. You asked it to make the button bigger, but it also changed the recording logic.

The way to deal with it is to make your words more precise: “The modification just now broke the recording function. Please keep the modification of making the button larger and repair the recording function back to its original appearance.” Telling it clearly which one should be kept and which one should be restored is much more effective than just saying “it is broken again”.

Another way to prevent this is to change one thing at a time. If you want to change three places, say it three times. After each change, confirm it before changing the next one. It seems slower, but actually saves more time.

Situation 2: The more changes are made, the more chaotic it becomes, and there is no going back

After the change, you no longer know which version is better.

In the web version, Claude can go back to a previous message, press Edit to resend, and the conversation will branch from there, which is equivalent to going back in time and starting over. Using the terminal’s Claude Code, you can directly say “Please restore the file to the last usable version.”

A better habit is: every time you complete a satisfactory stage, say to yourself “This is the save point.” The web version can save the reply, and Claude Code can ask it to make a version record for you.

Situation 3: The screen is blank, or an incomprehensible error pops up

At this time, there is no need to try to understand it yourself. Copy the entire error message on the screen and paste it to the AI, saying “This error occurred, please fix it and tell me the reason in one sentence.”

The last sentence “Tell me the reason in one sentence” is my little habit. You don’t need to become an engineer, but if you understand a little bit more each time, you will know better how to describe the problem next time.

Pick the first question and only look at one condition

The last piece of information is related to choosing a topic.

The first project that many people want to do is usually very large in scale: a complete e-commerce website, an app that can replace the company’s system. These are all very good, but as a first work, it is easy to fail on the third day.

For example, during a previous class, a student raised his hand and asked me: “Teacher, can I use AI to make Shopee?”

Therefore, I now suggest that you choose a topic based on only one condition:

Will you really open it next week?

It can be anything from keeping accounts, recording drinking water and sleep, helping clients calculate quotations, drawing lots for classes, attendance records for reading clubs, or making a month-end summary of fixed expenses at home… These topics are not big, but they have one thing in common: you will use them every day or every week.

Only when you know how to use it will you find out what doesn’t work well; when you find it doesn’t work well, you will want to change it; by changing it, you will develop your ability to express your needs. This cycle is more effective than watching any instructional video.

Doing it alone, the hardest thing is to survive the third day

At this point, you may have already figured out what to do. But I want to be honest about one thing: looking at this year, most people are stuck, but it is not technology.

Rather, a person only gives up after halfway through.

I was very excited on the first day and still remembered it on the second day. I put it aside when I was busy with work on the third day. It was already two months later when I opened it again and I had already forgotten where I had done it. I have received responses from many readers, and the stories are almost always like this.

So this time I want to try a new method and accompany you to complete the first work.

Seven-day challenge camp: Make your first AI gadget in seven days

The design of the challenge camp is simple:

• Time: November 30 (Monday) to December 6 (Sunday), 2026

• A task letter every day: sent to your mailbox in the morning. The daily task should be completed in 30 to 60 minutes.

• LINE learning group: Check in to the group after finishing every day. If you are stuck, just ask. I will reply once every night, and classmates can also help each other.

• The seventh day of acceptance live broadcast: At 8 pm on December 6, I will select several students’ works to watch live and discuss together how to make them more useful. Those who cannot attend can watch the replay later.

• Fee: NT$999, places are limited The seven-day progress is arranged like this: on the first day, choose a topic and let it get started; on the second day, use the five-stage formula introduced today to clearly explain the requirements; on the third day, use it in a real day; on the fourth day, make it easy to use on mobile phones; on the fifth day, find someone to try it out and see where he gets stuck; on the sixth day, publish it online and get your first URL; on the seventh day, the acceptance live broadcast.

After seven days, you will have a small tool that you use every day, put it on the Internet, have a URL, and can share it with friends.

Each task letter will be accompanied by two methods: those who use the claude.ai web version or Claude App do not need to install anything; those who are already using the terminal Claude Code can also follow the same progress.

One thing needs to be made clear first: Claude’s free account can do things, but the limit is very limited. If you do it for seven consecutive days, you will almost certainly hit the limit. Claude Code can only be used with paid plans. Therefore, I recommend subscribing to Claude Pro for one month (US$20) during the challenge period. The challenge camp itself does not pay the subscription fee.

Registration is paid through Portaly and electronic invoices can be issued.

See the complete introduction and registration of the Challenge Camp (opens in a new tab)

If you want to learn the entire method systematically, “Learning Vibe Coding from Zero Basics” covers everything from websites and apps to work automation. It has complete examples and debugging methods. It is also very suitable to continue reading after the challenge camp is over.

Get “Learning Vibe Coding from Zero Basics” (opens in a new tab)

If you want to warm up before starting the camp, you can read the free e-book “Your First Vibe Coding Project” to familiarize yourself with Claude’s operations. The first day of the camp will be much smoother.

Speaking clearly is an ability that can be practiced

Looking back on this year, I think what Vibe Coding (opens in a new tab) taught me the most is not how to make a website, but how to express a vague idea in a way that others can understand and do it.

To be honest, this matter is not unfamiliar to people who write. Before writing an article, you must first think clearly about who the readers are, what they want to say, and what they will not say; before asking AI to be a tool, you must think about the same set of questions. The only difference is that this time the person listening to you is AI, which will immediately turn what you say into something that can be pressed and used.

The five-part argument introduced today can be used immediately without signing up for Challenge Camp (opens in a new tab). Next time before you ask an AI to do anything, take five minutes to write it down and give it a try, I’m curious what you come up with? You are also welcome to reply directly and tell me, what is the first gadget you most want to make?

See you next time!

Vista

---

*This article is reprinted from “Vista Electronic News” No. 148〈AI can’t do what you want, probably because the requirements are not clearly stated (opens in a new tab)>, welcome Free newsletter subscription (opens in a new tab). *