When a Client Says 'Use AI to Build Me an Exact Copy of Notion, Budget 5000, Three Days to Launch' — What I'm Thinking

Foreword

Hi everyone, I'm Dreaming Cat, an independent developer.

I have a problem — I say yes to whatever a client asks. Not because I'm reliable, but purely because I'm thin-skinned and can't bring myself to refuse face-to-face. So over the years, the bizarre requirements I've taken on outnumber the README files on GitHub.

But since AI took off, I've discovered that the 'bizarre requirements' track has entered a whole new dimension. Before, when clients proposed requirements, they'd at least have a competitive analysis document — even if that document was just three screenshots. Now? A voice message gets tossed over on WeChat: 'You know AI, right? Write me an exact copy of Notion, every single feature, budget 5000, can go live in three days?'

I stared at the screen, hand hovering over the keyboard, thinking for a full three minutes — not about how to refuse, but about 'why did I ever tell him I know AI.'

Today, let's talk about a few uniquely AI-era bizarre requirements I've experienced firsthand, and how I dealt with them (painfully).

---

Requirement 1: 'Use AI to Write an Exact Copy of Notion, Budget 5000, Three Days to Launch'

Background

The client was introduced by a friend, someone doing cross-border e-commerce. On the phone, he said: 'I saw your Moments, you've been doing AI stuff lately, right? That ChatGPT thing can write code, can't it? Help me write a note-taking app, same features as Notion, for our team to use.'

I asked: 'What's the budget?'

He thought for a moment: '5000.'

Me: 'Timeline?'

Him: 'Three days, I need it urgently. It should be fast with AI, right? I read online that AI can write tens of thousands of lines of code a day.'

What I Was Thinking

My internal monologue went like this:

1. Notion is valued at 10 billion dollars, has over 300 engineers, and was developed over 7 years.
2. You want me to replicate it with 5000 yuan and three days.
3. And you think this is perfectly reasonable.

But I didn't say any of that. What I said was: 'Bro, Notion's block editor — just the cursor sync, drag-and-drop sorting, and nested indentation alone would take me a month to write.'

He said: 'Oh... can you simplify it a bit then?'

I said: 'Simplify how?'

He said: 'Just... remove the unimportant features, just leave the editor.'

Me: 'So isn't that just a Markdown editor?'

Him: 'Yes! That's about right!'

How I Handled It

In the end, I recommended an open-source solution — built a simple collaborative editor based on Tiptap, took about a week, charged 8000. Functionality-wise, it could only do basic rich-text editing and real-time sync, a hundred thousand miles from Notion.

But the client was very satisfied. Because what he wanted was never Notion; he just wanted a tool that 'allows multiple people to write documents at the same time.'

Lesson Learned: Often, when a client says 'make something exactly like XXX,' it's not because they're unrealistic, but because they don't know how complex that thing really is. They're just expressing a vague need — 'I want the kind of experience XXX provides.' What you need to do at this point is not complain, but translate what they actually need.

---

Requirement 2: 'Use AI to Write a Machine Learning Model to Predict Our Company's Stock Trend'

Background

This requirement came from a startup boss. On the phone, he said: 'I read those AI articles you wrote, seems like you know a lot about machine learning. Our company has a need — can you use AI to predict our company's stock trend?'

Me: 'Has your company... gone public yet?'

Him: 'Not yet.'

Me: 'Then how do you predict the stock trend?'

Him: 'Just predict how much it will rise after it goes public in the future.'

Me: '......'

What I Was Thinking

Everything I learned in four years of university flashed through my mind — linear algebra, probability theory, stochastic processes, time series analysis — and then I realized, none of these things can predict how much a company that hasn't even gone public yet will rise in the future.

But I didn't directly say 'that's impossible.' I know this type of boss hates hearing 'impossible' the most. I changed my phrasing:

'Boss, here's the thing: stock prediction falls under financial time series analysis. The mainstream models in this direction include LSTM, Transformer, temporal convolutional networks, etc. But the core problem is — we need historical data.'

He: 'So what do we do?'

Me: 'We can change the approach and do a competitive benchmarking analysis. I'll help you use AI to crawl some public data from listed companies in the same industry, make a comparison report, and see what dimensions your company has advantages in.'

He thought for a moment: 'Alright, that works too.'

The Final Solution

I spent about a week writing a competitive analysis tool in Python, crawling financial report data, social media sentiment, and recruitment info from several listed companies in the same industry, then built a simple comparative visualization dashboard.

Cost: Zero (except my time). Used public financial APIs and free NLP models for sentiment analysis.

The client was very satisfied, because that report helped him secure a round of financing. Although it was a hundred thousand miles from his original 'predict stock price' request, the result was far more useful than predicting stock prices.

Lesson Learned: When a client proposes a technically impossible requirement, don't refuse directly. You need to understand why they want that thing. This boss didn't want 'stock price prediction'; he wanted 'data to convince investors.' Once you figure that out, the solution presents itself.

---

Requirement 3: 'Add an AI to My Mini Program That Automatically Identifies What's in User-Uploaded Images'

Background

This requirement sounds pretty normal, right? 'Image recognition' — just call an API.

But the problem was — the client wanted to identify 'clothes in user-uploaded images, then automatically match them to identical items in our store.'

Me: 'This... how many clothing items are in your store?'

Him: 'About 20,000.'

Me: 'Do you have an image database?'

Him: 'No, they're all in our mini program, can you crawl them yourself?'

Me: '......'

What I Was Thinking

The 'identical clothing matching' requirement, seemingly simple, actually involves the following tech stacks:

1. Object Detection — Identify the clothing region from the image
2. Feature Extraction — Extract the clothing's color, texture, and style features
3. Similarity Retrieval — Find the most similar item among 20,000 products
4. Real-time Performance — Return results within 3 seconds of user upload

Any single one of these is a job for a large company team. And most critically — where does the training data come from?

How I Handled It

I spent three days doing one thing: downgrading the requirement.

I told the client: 'Identical matching can't be done in the short term, but I can make a simplified version — after the user uploads an image, the AI automatically identifies the clothing's color and type (e.g., "red dress"), then automatically searches the corresponding category in your store.'

He: 'How much worse is the result?'

Me: 'Identical matching accuracy might reach about 30%, but color + type recognition can exceed 90%. And the latter can go live in a month, while the former might take half a year.'

He chose the latter.

After launch, the results were pretty good: user uploads image → auto-identifies color and type → jumps to the corresponding category. Conversion rate increased by 15%.

Lesson Learned: For many requirements that seem 'AI can do it,' the core difficulty isn't in the AI, but in data and engineering. Rather than explaining concepts like 'feature vector retrieval,' 'vector databases,' and 'embeddings' to the client, it's better to directly offer an 80-point alternative solution. An 80-point solution can go live; a 100-point solution can only live in a PowerPoint.

---

Requirement 4: 'Make Me an AI Customer Service That Can Automatically Reply to All Customer Questions, But Can't Let Customers Discover It's AI'

Background

This requirement came from a friend doing e-commerce. He said: 'My customer service team is too expensive, costing twenty to thirty thousand a month. Can you use AI to make me a customer service bot that automatically replies to all questions, and can't let customers discover the other side is AI?'

I said: 'Isn't this... asking a bit much?'

He: 'Not much, I just have two requirements:

1. Can answer all customer questions
2. Can't let customers discover it's AI
3. Budget 3000'

Me: 'That's three requirements. And 1 and 2 are contradictory — an AI that can answer all questions, customers will definitely discover it's AI.'

He: 'Why?'

What I Was Thinking

Customer: Do your clothes shrink? AI Customer Service: Our clothes are produced strictly according to national standards, with a washing shrinkage rate controlled within 3%, compliant with GB/T 22849-2014 standard. Customer: ...Are you a robot?

I've rehearsed this conversation in my head 100 times. The biggest problem with AI customer service isn't that it's not smart enough, but that it's too smart. A real person would say 'No way haha, our clothes are great quality, I wear them myself,' while an AI customer service would say 'According to our quality control standards...'

The Final Solution

I made him a 'hybrid customer service' solution:

1. AI handles common questions (logistics tracking, return/exchange rules, size recommendations) — covering 80% of issues
2. Complex questions automatically transfer to human agents — when the AI judges it might not handle it well, it automatically transfers to a real customer service rep
3. AI reply style adjustment — added a colloquial vocabulary, occasionally deliberately makes typos, uses emojis, making replies seem more human

After launch, customer service labor costs dropped by 40%, and because the AI handled most repetitive questions, the job satisfaction of real customer service staff actually increased.

Lesson Learned: Sometimes, 'making AI seem human' is more important than 'making AI smarter.' Users don't care how many parameters your model has; they only care — 'is this person brushing me off when they reply?'

---

Requirement 5: 'I Want to Make an AI That Can Automatically Write Juejin Articles for Me, One Per Week, That Can Hit Trending'

Background

This requirement... I gave it to myself.

Yes, I actually tried this. I wrote a script that used AI to automatically search Juejin trending topics, analyze the characteristics of trending articles, generate articles, and auto-publish. At the time, I thought this thing was absolutely perfect — gain followers while lying down.

Then I published the first one.

Reads: 47. Likes: 0. Comments: 0. Saves: 0.

I stared at the backend data for ten minutes, then silently went to read that article.

Honestly, the writing wasn't bad — clear structure, logical flow, even had code examples. But the problem was also obvious: no human touch. Every sentence was like a textbook, every example like documentation. After reading the whole article, you had absolutely no idea who 'this person' was, what they had experienced, why they wrote this article.

My Reflection

Later, I summarized several fatal problems with AI writing articles:

1. No personal experience — AI can say 'Vue 3's reactivity principle is based on Proxy,' but it can't say 'The first time I used Vue 3, I debugged for a whole day because I mixed up reactive and ref'
2. No real emotion — AI won't say 'This requirement made my blood pressure spike,' it will only say 'This requirement presents certain technical challenges'
3. No opinion — AI only lists facts, it won't say 'I think this solution is better than that one, because...'

So now, my method for using AI to write articles is: Let AI help me look up information, organize structure, and polish language, but the core opinions and personal experiences must be written by myself. AI is a tool, not a stand-in.

---

Final Words: Bizarre Requirements in the AI Era Are Essentially a Cognitive Gap

Looking back at these requirements, I've discovered a pattern:

Clients propose bizarre requirements not because they're stupid, but because they have a massive cognitive bias about 'what AI can do.'

They see news saying 'AI writes tens of thousands of lines of code a day' and think AI can write a Taobao in a day. They see 'AI customer service can auto-reply' and think AI can perfectly replace human customer service. They see 'AI predicts stocks' and think AI can predict anything without historical data.

And our job is not to mock these requirements, but to translate the gray area between 'what AI can do' and 'what AI cannot do' into language the client can understand.

This is much harder than writing code. But it's also the core value of us tech people in the AI era — not whether we can use AI, but whether we know when to use AI and when not to.

What bizarre AI-era requirements have you encountered? Welcome to share your stories in the comments, I really want to see everyone's 'blood pressure spiking moments' 😂

---

#BizarreRequirementsAwards #AI #ArtificialIntelligence #IndieDeveloper #ProgrammerLife #ProductManager