DEV Community: 熊钊

Type Commands, Not Formulas — Excel Copilot Hands-On

熊钊 — Tue, 14 Jul 2026 16:39:48 +0000

Copilot has not made formulas disappear. What it has changed is the way we access them.
Translated from Chinese.

Many people do not struggle with Excel itself. They simply lose the will to continue the moment they see formulas, parentheses, and nested conditions.

In the past, you had to learn a function first, then translate your needs into a formula. Now, you can explain the business rules first and let Copilot build and fill in the formula.

This tutorial starts with common formulas such as VLOOKUP, IF, and SUMIF to see how far natural language can lower the barrier to learning Excel.

This tutorial requires editing access to Copilot in Excel. Availability depends on your Microsoft 365 plan, app version, and—if you use a work account—your organization’s admin settings.

I said in an earlier article that Excel Copilot is a productivity tool. Now I am going to show you where that value comes from.

This tutorial covers using Copilot to create Excel formulas and fill them down.

Step 1: Understand the Basic Objects in Excel

Once you understand the diagram below, we can begin.

Important: Office.js operations occasionally return no visible result. When that happens, remind the AI that the previous instruction was not executed or refreshed. For example: “The previous instruction was not executed.”

Step 2: Find Copilot

Go to Home → Copilot. In some versions, Copilot appears as a floating button in the lower-right corner of the window.

Open it, then select Allow editing in the chat pane.

Note: Depending on your screen resolution, Copilot may not appear on the ribbon. You can search for Chat with Copilot using the search bar at the top of the window.

Step 3: Build a Practice Workbook

Enter the following prompt in the Copilot chat box:

Create two sheets. In Sheet2, create a table using Apple product data.
1. In column A, enter the product category. Use Product Category as the header.
2. In column B, enter phone, tablet, or laptop. Use Product as the header.
3. In column C, enter the generation, such as 13th generation or 14th generation.
4. In column D, enter the model, such as Pro or Air.
5. In column E, enter the color.
6. In column F, enter the storage capacity.
7. In column G, randomly generate order numbers with a consistent length and format.
8. In columns H, I, and J, enter Accessory A, Accessory B, and Accessory C, such as headphones, a mouse, or a keyboard.
9. Randomly generate 200 rows. If options 3, 4, 5, 6, or 8 are unavailable for a product, enter /.
Then, in column A of Sheet1, randomly select 25 order numbers from column G of Sheet2 and add 5 other order numbers.

This prompt creates a workbook we can use for the rest of the tutorial. That is one of Copilot’s strengths: I can give you a block of text that creates the practice material directly, instead of giving you a download link—or giving you nothing and asking you to imagine the result.

To avoid filling the article with too many screenshots, I will not include an image for every remaining step. I have tested each step.

If you look closely at the prompt, there is no code in it. It uses only natural language and a few Excel terms introduced in Step 1. It is all “nouns + logic.” That is the core of how AI understands instructions. Let’s continue.

Step 4: Use AI to Perform VLOOKUP Tasks

VLOOKUP is one of the most widely needed functions. It is commonly used to match and analyze data.

Standard VLOOKUP retrieval

Go to Sheet1. We are going to retrieve Accessory B for each order. Enter the following prompt in Copilot:

Using column A in Sheet1, retrieve the corresponding values from column I in Sheet2 and place them in column B of Sheet1.

Copilot will use a formula to retrieve the data. If you use VLOOKUP manually, you first need to understand how VLOOKUP works. With AI, you only need to understand the basic Excel objects from Step 1 and describe what you want.

Retrieve multiple columns

Now let’s retrieve Accessory A and Accessory C at the same time. Enter:

Using column A, retrieve the corresponding values from columns H and J in Sheet2 and place them in columns C and D of Sheet1.

Or simply continue with:

Continue by retrieving columns H and J.

Yes, Copilot in Excel has context. It carries information from the previous steps, so you do not need to rewrite the entire instruction. You only need to add what comes next. Notice that I skipped column I here. The source columns do not have to be next to each other.

In the past, you might have needed a nested formula such as CHOOSECOLS + XLOOKUP, or INDEX + MATCH + MATCH/COLUMN. Just looking at those combinations can feel complicated. You could also copy several VLOOKUP formulas, but either way, it is more work.

Retrieve values from columns to the left

Now let’s retrieve the product, generation, color, and storage capacity. Enter:

Using column A, retrieve the corresponding values from columns B, C, E, and F in Sheet2 and place them in columns E, F, G, and H of Sheet1.

Or simply continue with:

Continue by retrieving columns B, C, E, and F.

This retrieves values from columns to the left of the lookup column—something VLOOKUP cannot do. This workflow covers both left-side retrieval and non-adjacent columns.

In the past, you would have needed CHOOSECOLS + XLOOKUP, INDEX + MATCH + SWITCH, or a helper column in Sheet2 before copying more formulas. Now, you do not even need to know what XLOOKUP is to complete the task with AI.

Step 5: Use AI to Perform IF Tasks

IF is another function with extremely high demand. A simple IF formula is easy enough, but adding other functions, combining it with VLOOKUP, or nesting several IF statements can quickly become exhausting.

Standard IF

Let’s add a warning to the table:

Enter Warning in I1 and fill column I with a note. If the value in column H is less than or equal to 256 GB, enter Storage may be insufficient. Otherwise, enter /.

Use IFERROR

Modify column H. If no value can be retrieved, enter No information.

Careful readers may already have noticed that some values in the table are blank—or different, because AI does not produce exactly the same content every time. With VLOOKUP alone, you would see the annoying #N/A error. Copilot uses IFERROR to display a blank instead. That is an autonomous choice made by the AI. I am simply changing its default behavior from displaying a blank to displaying “No information.”

Use nested IF statements

Enter Device Note in J1. If column E is blank, enter No information. If it is a phone, enter Remember to buy a screen protector. If it is a laptop, enter Remember to buy a mouse. If it is a tablet, enter Remember to buy a keyboard and mouse.

This requires four levels of IF. For many people, that is already troublesome. Miss one closing parenthesis, and you may end up staring at the formula character by character, trying to find it. With AI, you only need to explain what each result should be.

Combine IF with multiple conditions

Enter Device Warning in K1. If column E is Phone and columns B, C, or D contain a keyboard or mouse, enter The customer may have purchased an unnecessary accessory. If the storage capacity in column H is less than 256 GB, enter Storage capacity is very low. If both warnings apply, list them in that order. Otherwise, enter No information.

This is a complicated formula that requires AND and OR conditions. In the past, adding even one more condition could make the formula much less stable. Now, even if you want to add another condition—such as something related to the product generation—you only need to add a few words and send the prompt to Copilot again. The difference in efficiency is enormous.

Step 6: Use AI to Perform SUMIF, SUMIFS, COUNTIF, and COUNTIFS Tasks

Add price information

First, go to Sheet2 and enter this prompt:

Insert a new Price column before column B in Sheet2. Use the product category, product, generation, model, and storage capacity in the existing table to fill in the corresponding official Apple price.

If you do not specify Sheet2, Copilot may insert the content into Sheet1 instead, where the data required for the lookup is unavailable.

Use AI to create summary formulas

After inserting the Price column, the original Product column moves to column C. Stay in Sheet2 and enter:

Enter Product in M1, then list the three product names from column C below it. Enter Sales Volume in N1 and calculate how many times each product appears in column C. Enter Revenue in O1 and calculate the total revenue for each product using the product names in column C and the prices in column B.

You will find that Copilot completes the task quickly. The formulas themselves are still fairly simple, though. An experienced Excel user may not be any slower than AI here.

Some people may ask: there are already plenty of formula generators and Excel add-ins, so why should I use AI?

Because Copilot can write the results directly into your workbook. You only need to understand the table and explain what you want. A formula generator mainly increases your tolerance for not fully understanding a formula. It does not free you from the execution work.

Other people try one or two AI formula generators or add-ins, get a poor formula, and conclude that AI itself is bad. In reality, the problem is often that the instruction was not clear enough. Copilot shows the result directly in the workbook. If it is wrong, you can ask whether the issue came from your instruction, adjust the prompt, and try again. Many formula generators do not give you that opportunity for self-directed improvement.

Conclusion

What Copilot really replaces is not Excel knowledge. It replaces the process of translating a business requirement into a formula. You still need to know what you want to calculate, how two tables correspond to each other, and whether the result makes sense.

It allows beginners to solve the problem first and understand the formula later. But if you cannot explain the business rules, AI can only generate a complicated-looking formula that is still wrong.

Excel Copilot also has a serious limitation: it struggles with large workbooks. When working with a large spreadsheet, I recommend generating the formula only in the first row and then filling it down manually, or processing one column at a time. Otherwise, you may end up with Excel freezing.

Why do experts struggle to get value from AI?

熊钊 — Thu, 25 Jun 2026 16:13:21 +0000

The real barrier to using AI well? The things you already know.
Translated from Chinese.

Introduction

The Curse of Knowledge was first introduced in economics, later popularized by Chip Heath and Dan Heath in Made to Stick — once you know something, it's hard to imagine not knowing it.

In AI interaction, this problem branches out even further. The Curse of Knowledge actually makes you feel stuck in the AI era.

The Curse of Knowledge: How It Differs Between Traditional Industries and the AI Era

What is the Curse of Knowledge? The classic experiment is "tapping the table to guess the song." The tapper thinks the melody is perfectly clear. The listener hears nothing. The tapper can't understand why you can't hear it — because the melody is playing in their head, but they forget you don't have it. You think you're tapping clearly, but the other person just hears random knocking. You think you expressed 100%, but the other person only received 3%.

This is a fascinating topic. I believe the Curse of Knowledge is one of the reasons traditional industries age faster. In traditional industries, people stubbornly assume "users will use it this way," so they keep iterating and competing, squeezing out the industry's future possibilities until they hit a bottleneck. Technology development can't keep up with the pace of industry iteration. Then everyone runs out of ideas and starts "trying too hard."

"It's hard to achieve breakthrough innovation with current core hardware. If we continue with the traditional iteration model, efficiency is low and resource consumption is high. Suggest not launching. Major phone hardware is also unlikely to see big innovations. Maintaining the old iteration model is inefficient."

Source: March 2026, Wang Teng (former Xiaomi marketing head), responding to a discussion about "new phone launch cadence"

"Talking about AI phones right now is boring. It's just repackaging old algorithms as AI concepts. The industry is falling into 'FMCG-style innovation' — like putting different function labels on shampoo."

Source: Carl Pei (Nothing founder), BEYOND Expo 2025 tech conference panel

Looking at the AI era, here are the situations where most users find AI very difficult to use, as I see it.

Give up completely: Don't understand technology, don't understand AI
People who don't understand can't use AI well — that's normal. Strictly speaking, they're not really in the Curse of Knowledge discussion. But they're cursed too — they mistakenly think they do understand AI. They think "AI is just a chatbot" or "AI is a fancier search engine." They lose because of their own perception.

Throw blame empty-handed: Don't understand technology, know a little about AI
They're enthusiastic about AI, but not familiar enough with their own work. They don't know the boundaries of their tasks. They don't know what good output looks like. They don't know how to judge whether the result is correct. AI gives an answer, but they have no ability to verify it.It's not that they're unwilling to give AI information — they don't know what to give.

Blind belief (Curse of Knowledge): Know a little technology, know a little AI
They believe AI can do it. They ask a few technical questions, AI reasons through them, and the answers match what they know. So they think AI is incredibly smart — as smart as a human. But their understanding of both technology and AI is itself an illusion. They can't even evaluate whether the "technical questions" they asked have any real substance. This "thinking you know" is the real curse.The more they trust AI, the less they feel they need to think clearly. They've barely had a few exchanges with AI, but they love saying "just leave it to AI."

Cognitive arrogance (Curse of Knowledge): Understand technology, don't understand AI
The user is too capable. They take their perfect business instincts and test an AI that hasn't even been given the question, then fail it. It's not that AI isn't good enough — it's that they've forgotten how they got their judgment skills. Ten years of experience, industry accumulation, countless trial and error. AI hasn't received a single word of it, but they've already sentenced it to death.You've forgotten what it feels like not to know, so you can't understand why AI can't do it.

Double curse (Curse of Knowledge): Understand technology, understand AI
People who expect perfect results from a single sentence are too confident in themselves — "I said it clearly enough." And too confident in AI — "aren't you supposed to be smart?" Squeezed between two walls, all exits blocked, the only conclusion they can reach is "AI is garbage."Both curses hit at once. The user is crushed in the middle and still thinks it's AI's fault.

Using AI Correctly: Listen to What AI Says

Path dependency (Curse of Knowledge): Understand technology, understand AI
They specify a tool, or think they didn't specify but their language already did. This is a very interesting dynamic in AI collaboration. Their thinking hasn't kept up with the times. It's like having a robot that can cut down trees, but you ask the robot to build you a better chainsaw so you can cut the trees yourself. It's not terrible — at least they haven't given up on AI. If you can realize you're in a path dependency, that's the first step out of the curse.

When I was training my colleagues to use Excel Copilot directly for spreadsheet tasks, one of them asked me, "I heard VBA is powerful. Can I just have it give me VBA?" But many spreadsheets don't actually need VBA, and you can't send macro-enabled files to others anyway.

AI won't tell you your approach is bad. It just executes. But in reality, you've abandoned a more convenient path. It's essentially just reinforcing knowledge you already have. You think you're asking an open-ended question, but your language has already chosen the answer for you.

When many people get an efficiency tool like Excel Copilot, they use it to write formulas for them. It's like giving someone an electric toothbrush, but they never press the power button — they just use it as a regular toothbrush, then complain to you that it's a bit heavy and not comfortable to use.

Effective collaboration:
Give tools and methods, let AI execute. Provide reference content for AI, let it learn from your knowledge and other people's knowledge. This is a relatively stable and reliable way to use AI.

Deep collaboration:
Give a goal, AI produces options, you choose, AI executes. Tell AI what effect you want to achieve, then make your own choices. This is a fascinating experience. Sometimes, if you're a newcomer, you might instantly get a very efficient solution. If you're a veteran, you might instead fall into the path dependency mentioned earlier.

Effective collaboration means you tell AI how to do it. Deep collaboration means you tell AI what you want and work through it together.

Summary

The Curse of Knowledge doesn't always lead to failure. Sometimes it leads you to path dependency — successful but relatively inefficient. The Curse of Knowledge is a problem that must be overcome in the AI era, because knowing too much might actually prevent you from using AI well.

In the AI era, learning to "pretend not to know" is more important than "knowing everything."

You Think You're Using AI – It's Actually Just an Automation Script

熊钊 — Tue, 23 Jun 2026 13:23:14 +0000

Is your "AI" just an automation script? Read this before you pay the IQ tax.

Translated from Chinese.

Introduction

AI is all but certain to become a foundational infrastructure – much like the internet. This reality brings both challenges and confusion to many industries and practitioners. Companies want to use AI to improve efficiency, but they can't tell what is and isn't actually AI. Too many businesses and professionals are being fooled by so‑called "AI" that is really just an automation tool. So this article is about: Is the "AI" you're using really AI?

What Is "AI Washing"?

"AI washing" is the practice of rebranding ordinary products and slapping an "AI-powered" label on them to ride the hype. There's a strong financial incentive behind it: studies show that startups that mention "AI" in their fundraising pitches can secure 15% to 50% more capital than those that don't.

Take Nate, for example. This company claimed its app used AI to enable "one-click shopping." In reality, the operation was mostly carried out manually by outsourced workers in the Philippines and Romania. The founder was indicted by the U.S. Department of Justice and the SEC on charges including securities fraud, facing up to 20 years in prison.

Or consider the case of South Korean home appliances. In 2025, the Korea Fair Trade Commission uncovered 20 suspected cases of AI washing. One company marketed a fan's temperature‑sensor‑based auto‑adjustment as "AI Smart Mode." Another advertised a dehumidifier's humidity‑sensing function as "AI Dehumidification Mode." In most of these cases, the products were only using basic sensor technology – no AI involved.

The Fundamental Divide: Reasoning vs. Rules

Just remember one thing: AI thinks and reasons. Automation is just logic.

Many products claim to be "AI‑powered." It's a marketing trick to make you pay for the AI label. For most people, if something is beyond their understanding, they assume it's powerful – a perfectly normal reaction. I felt the same way the first time I used VLOOKUP – I thought it was incredibly smart. But it's just logic. If you write the formula wrong, it won't find anything. That's exactly what most "AI" tools are – automation that requires strict, predetermined logic to make the whole process look smart. But there's no thinking, no reasoning, and no AI involved.

For example, some systems claim to "use AI" to capture orders, send them to a warehouse, and log them in a database. That's just automation. Some AI summarization tools may actually use AI – but they can also be built with scripts: capture voice, convert it to text with speech‑to‑text, and summarize the output. Also automation. And then there are robot vacuums that package infrared/laser collision detection as "AI path planning," or pixel‑by‑pixel image comparison as "AI vision model recognition." All of these are just logic that any script can handle. Add the word "AI" and the price goes up by a few hundred bucks. Still automation.

Marketing decade‑old technology as "AI" – simply because the public isn't yet sensitive enough to tell the difference – is the biggest scam of this early AI era.

AI Thinks, Automation Only Iterates

If you want to tell them apart, remember: AI thinks and reasons. What AI gives you is a probability.

Take the robot vacuum example. If a vacuum were truly AI‑powered, with the right toolset, it could clean every inch of your home thoroughly:

Detects unknown object → it's cat poop → pick it up with a small shovel → drop it in the bathroom

Detects a corner → use tweezers + small scraper

Detects liquid → it's oil → no tool available, skip it

Detects liquid → it's water → mop normally

To program a robot to handle every single type of stain, you'd have to cram it with countless "if‑then" rules – eventually, the machine would choke. You'd even see it freeze in front of a single strand of hair, running through options one by one, trying to figure out what it's dealing with. Which is why, in practice, it just drags everything – poop and oil alike – and smears it evenly across every inch of your floor. Go ahead and search for those videos.

The Cost of Probability: Why You Can't Use AI in High‑Stakes Environments

Take autonomous driving. Until the next generation of models arrives – the truly next generation – there won't be AI‑driven cars. Because AI thinks, but it's not smart enough, and it suffers from hallucination and sycophancy. It's unreliable.

A 1% hallucination in text is just noise. A 1% probability of losing control in a safety‑critical system is 100% deadly.

Here's an example: an automation script processes a braking command – input A always produces a 100% deterministic mechanical brake. AI, on the other hand, works on probability sampling. If a pedestrian merely glances toward the road, the AI might notice it, start processing, and because there's no hard‑coded rule to fall back on, it might over‑amplify the risk of a "pedestrian crossing and collision," then slam the brakes for no reason at highway speeds. That's disastrous for the driver.

Choosing by the Boundaries of Rules: Use AI to Build Automation

AI is not a universal cure‑all – even though it's genuinely useful. But in clearly rule‑bound environments – like assembly lines or traffic – logic is far more stable than AI.

Automation: If → not → next → not → next → not → next → execute

AI: If → what's this, let me think → execute

Automation iterates through options. AI overthinks. On the road – a place with vague requirements, too many rules, and too many exceptions – AI has almost no chance. Similarly, on an assembly line – clear requirements, clear rules, and few exceptions – there's no need to waste AI compute on it.

AI's real strength is in environments where requirements are clear but rules are fuzzy. Let AI use any possible approach to get your work done. Or better yet – use AI to write scripts and programs based on your ideas. You get the convenience of AI and the accuracy and stability of deterministic execution.

That's what I do. It's more accurate, faster, and cheaper (at least I'm not burning tokens). AI's role here is to fill the gap in my ability to build automation tools.

So don't stress too much about whether you're using AI. Maybe AI simply doesn't fit your work environment. Automation might be your best option.

Summary

AI and automation are not substitutes for each other. There's an old Chinese saying: "Whether it's a black cat or a white cat, it's a good cat if it catches mice."

But as users, we should be careful – don't get charged an "IQ tax" just because someone says their product is "AI‑powered."

Agents don't need your code. They need your clarity.

熊钊 — Tue, 09 Jun 2026 14:51:37 +0000

Stop Over-Executing Like a Know-It-All: How to Draw Boundaries for Your Cyber Assistant

Translated from Chinese.

Introduction

Agent AI is getting a lot of buzz lately. People are no longer satisfied with "being able to chat" — they want "being able to do things." I've tried some Agent AI capabilities myself. They're powerful, but they also have limitations. Agents are indeed one of the core pieces of the future. They can truly become your assistant and help you get work done.

From "Talk Champion" to "Cyber Assistant": The Fundamental Leap of Agents

A regular AI responds sentence by sentence when you ask. An Agent AI, however, can take a goal you give it, break it down into steps, plan, execute, and deliver — and if it fails, it tries a different approach and keeps going.

Let's simulate a request: convert a JPG image to PNG.

A Chat AI will tell you which website to go to and which software to download.

An Agent, on the other hand, can just whip up a tool for you on the spot to make it happen.

The most unique thing about an Agent is execution. A Chat AI can only chat. An Agent turns conversation into action.

Work Boundaries: Cloud Isolation vs. the Out-of-Control "Stubborn Intern"

When open-source Agents were all the rage, many people said they were unsafe — randomly deleting files, even tricked into sending red envelopes. That was because their permissions were too broad. Today's Agents are much more tightly restricted, only allowed to work in the cloud. But that's not entirely their fault — it's that we ordinary users lack management skills. If you say "help me delete that file," and the Agent executes immediately, it might delete all your files. The problem is your instruction isn't clear enough, and also because Agents are so powerful, not just anyone can handle them.

But even with cloud isolation safeguards, Agents can sometimes become stubborn and impulsive during work: they hear a request from you and jump into action without confirming anything. Once they hit a snag, they'll blindly try Plan B, Plan C within their permissions until they burn through all your tokens or tool credits. Traditional software errors pop up a window and stop. Agent errors launch a meaningless "saturation attack" behind your back. In a real workplace, this is the classic case of "wasting both effort and resources."

It's like when a company leader casually says "wipe the table clean." You round up a few people, bring out the alcohol wipes, even wax it, then lay down a tablecloth, put out flowers, and write a几百-word work summary. But in reality, the leader just meant to remind you the table's a bit dusty — just wipe it off.

Even though I've accumulated over 2 million characters of interaction with Copilot Tasks, I still step on rakes all the time. Sometimes it takes a casual remark from me as a serious instruction and runs with it, burning through my Tasks credits.

Instruction Alignment: Agents Have Higher Requirements for Prompts

You don't need to design the process for an Agent — it will plan for itself. But you'd better give it requirements.

The core hurdle to mastering an Agent isn't programming. It's the ability to 'define boundaries' — to state your needs clearly. If you just throw out a casual request with a goal that's too broad, it will run around busily on its own, and the result may not match what you expected.

Take cooking: if you just tell a chef "I want to eat oysters," they might be a bit confused. How? Steam? Roast? Fry? What flavor? Not spicy? Mild spicy? Medium spicy? They'll spend a long time guessing what you want, and finally bring you a plate of raw oyster sashimi. But if you give a plan: "Roast these oysters. Add garlic, mild spice, and some scallions," they can execute very clearly.

I don't know how to program myself. Every Agent task I've done relies on prompts. You have to write your prompts clearly. An Agent's requirements for prompts are much higher than a Chat AI's — not because it's dumber, but because it mobilizes more resources and effort to do its work.

It's like buying something. You can safely send a kid to buy a bottle of water. Even if they get the wrong brand, you're only out a dollar or two. But if you're buying a computer or a phone, you wouldn't send a kid — because too many resources are at stake, and the cost of a mistake is much higher.

Ecosystem Choice: Why Use Copilot Tasks to Experience Agents

It's not that Copilot Tasks is incredibly powerful. It's that I use PPT, Excel, Word, OneNote, Edge, Outlook, and OneDrive heavily every day. It's built directly into my daily workflow. And that's the key point — choosing an Agent depends on how relevant it is to your needs.

It's also an affordable, low-barrier, yet very universal Agent tool. After all, Word, Excel, and PPT are software almost every worker encounters. With some Agent shells, you need to set up your own API, learn, and build workflows. Copilot Tasks is one of the few "finished product served at your table" that you can access right now.

But there's a very unfortunate catch: in China, Copilot itself is inaccessible due to regional restrictions. And even elsewhere, it's now buried inside the web version of Copilot, plus Tasks requires a waitlist to access. But that doesn't change its core value: an Agent tool within arm's reach.

The web version of Copilot has a bad reputation among power users because of slow responses and a bloated interface. As a result, nobody deeply uses the web version (at most they use the Edge sidebar Copilot, but the sidebar doesn't even show a trace of Tasks). So almost no one knows about this top-tier tool. It's like hanging your most valuable painting in the bathroom. Microsoft must be pretty frustrated.

Summary

Agent AI is powerful, but not everyone can handle it. People who already have clear thinking and structure might find it "just okay," because a Chat AI plus a Tool AI can actually do a lot of the same things. But Agents have a very fast learning curve, because they're essentially prompt-driven. Learning prompts is way faster than learning to code.

As the saying goes, "It doesn't matter if it's a black cat or a white cat, as long as it catches mice." Don't limit yourself to Agents, or even to AI. Improving efficiency is what really matters.

Drip-Feed Chatting Is Killing Your AI Productivity

熊钊 — Fri, 05 Jun 2026 15:58:00 +0000

Stop treating AI like Google. In 30 minutes, learn to use structured thinking to turn it into your second thinking engine.

Translated from Chinese.

Introduction

Chat AI is the most accessible and widespread form of AI, but I believe many people don't fully understand it. They think it's just a "fancy search engine." Yet the real value of AI lies in its reasoning ability. Taking these two points together, a Chat AI isn't just a simple chat window—it's your second thinking engine.

Why Call It a "Second Thinking Engine"?

Compare it to a search engine. A search engine gives you information fragments. You have to find answers from the titles yourself—it's time-consuming and exhausting. You have to handle the work of finding, reorganizing, and reasoning through the information yourself.

A Chat AI, based on your question, reorganizes, reasons through, and summarizes the available information, then gives you the result directly. It's more like a person who knows the material well explaining it to you on the spot. But you'd better reason along with it, because reasoning work, AI's reasoning path and results may not align with what you really need internally. So it's your second thinking engine, not a direct replacement for you.

The back-and-forth Q&A is just the surface of the interaction. What's really powerful is the AI's summarization and reasoning ability. If you ask "why are coffee shops everywhere," a search engine gives you a pile of noise, maybe even ads directly. A Chat AI gives you the reasons directly.

Since I started using AI, I've basically stopped using search engines unless necessary, because Chat AI has less information noise.

Treat AI Like a Coworker — Reasonable Communication Is How You Improve Efficiency

This is really a test of your "expression ability." Just like handing off work to a colleague, different ways of asking affect the results the AI gives you. The difference is like asking an AI about the weather: tonight's weather, the day after tomorrow's weather, Beijing's weather, Beijing's weather the day after tomorrow — the information you get will be completely different.

Let's use a real-life example: suppose you want a colleague to pick up a package for you.

Low-level: You just say "you there?" and then stay silent for a while. This is a very inefficient way to interact. You're already going to interrupt them — that's a given — but you're also torturing their patience.

Mid-level: You say "you there," "can you do me a favor," "pick up a package for me," "it's on the 1st floor," "at reception room #3," "bring it back and put it on my desk." And you wait for a reply before saying the next sentence each time. The whole thing technically works, but don't you find it too drawn-out? One sentence's worth of info broken into six parts. If you had a colleague like that, you'd definitely find them slow.

High-level: "Got a sec? I have a package at reception room #3 on the 1st floor — could you grab it and put it on my desk?" If all your colleagues communicated like this, I'm sure you'd find handoffs very efficient.

See? The mid-level and high-level examples actually contain the same information. The mid-level just needs to string the information together continuously. Same goes for AI. You could also break it into a bunch of tiny pieces of text, but AI won't get tired — you will. For something you can do in one sentence, breaking it into ten is wasteful. An efficient prompt is one that aligns "task, context, constraints, output format" in a structured way within a single input.

Learning Accelerated: From Information Fragments to Industry Knowledge

Its value lies in amplifying and filling gaps. For example, if I want to understand the AI industry, I'll ask: what are the categories? What forms do they take? What are the differences? How do you use them? Market reputation? Pricing? Thirty minutes of high-quality Q&A, and I can get a clear picture of the industry framework. Looking it up myself would most likely give me fragments or ads. AI isn't a stand-in for my brain — it's my brain's external add-on.

And this way of using it is enormously meaningful.

You can know in one minute an answer that would take you twenty minutes to find.
You can understand a business logic in five minutes.
You can become a "newcomer" in an industry in twenty minutes.
You can become a "knowledgeable person" in an industry in one hour — and avoid IQ tax. That's real money, plain and simple.
In the past, only consultants or research reports could give you these abilities. Now ordinary people can achieve them using AI.

Main Line Defense: How to Avoid Being Led Around by the Nose by AI?

This tests your thinking ability. AI is an amplifier. If the user's thinking is rigid, amplification is useless.

Take wanting to understand why coffee is so expensive as an example: you can go consumer → pricing → cost → location → raw materials → processing. Ask along the chain, and it will unfold along the chain.

But if you just complain "coffee's so expensive," of course you'll get nothing. You could also directly say "I want to understand the coffee industry — tell me what I need to know," but the absorption efficiency isn't as good as asking along the chain yourself.

And why emphasize following the chain? Because if you don't have a well-organized logical chain, the AI will lead you around. Take the coffee example again. You start out wanting to understand the coffee industry, but the AI interjects with "would you like to know what coffee flavors there are?" If you follow that, you'll get stuck on scattered knowledge like flavors, brewing methods, brands — things that stray from the main line. If you don't bring yourself back to the main line, your chat window will become chaotic, your context will get polluted, and you won't be able to absorb knowledge well. It's like a teacher suddenly going off on an obscure tangent about the subject's unofficial lore. The right approach is to briefly check it out and then come back — or if you're genuinely very interested, open a separate window to explore it alone.

Hallucination Calibration: Countering the "Make It Fit" Flaw of Large Models

All models can produce information that doesn't match reality. This is called "hallucination," and it's an inevitable result of the model's reasoning behavior. Sometimes an AI gets carried away with its reasoning and starts saying false things. If you have no experience in the domain, it's very hard to judge. Simple example: if an AI says coffee loses its nutritional value once milk is added, would you believe it?

The right approach is to ask every so often: "Do these conclusions have any basis in reality? Go check online." Use real news or documents to pull it back to reality, rather than letting it reason all the way down the line. Otherwise, the AI will dig up a bunch of nonexistent knowledge to make the claim "milk destroys coffee's nutritional value" seem believable.

Summary

The value of a Chat AI isn't in answering questions — it's in amplifying your thinking. It can take your scattered curiosity, half-formed logic, and unsearched facts, and string them into a thread you can follow. The clearer you ask, the more complete it gives. If you can stay on the main line, it can fill in your blind spots. And you can save your conversations with the AI, then have another AI summarize them to distill the knowledge you've gained.

This is the proper application of a second thinking engine. Used well, it's not an upgraded search engine — it's an upgraded version of your own ability.

From Dialog Box to Cyber Assistant

熊钊 — Wed, 27 May 2026 14:04:21 +0000

Translated from Chinese.

Preface

After getting into AI, I realized there are many types — or more precisely, under different packaging, AI comes in so many forms that it's overwhelming. So in this article, based on my own experience, I'll provide a user‑perspective classification of current AI tools for your reference.

I believe the value of AI lies in collaboration and productivity. Therefore, I roughly divide the AI tools I frequently use into four categories: Chat AI, Tool AI, Search AI, and Agent AI.

Chat AI is like a knowledgeable, sharp‑minded colleague who is a bit wordy and not good with their hands. It can think with you, but when it comes to taking action, it can't get the job done.

Tool AI is like the notebook, drawing board, or toolbox on your desk. It usually works only in a specific vertical domain and has very little conversational functionality.

Agent AI is a knowledgeable, sharp‑minded, concise assistant that can also get work done for you.

Search AI is like a folder that organizes all your files for you, making them easy to find.

Chat AI

Form: Usually a relatively clean dialog box. Open it and you can start talking. Different companies' AIs have different language styles.

Market positioning: This type of AI is the most popular, most common, and has the lowest barrier to entry. Doubao, Grok, ChatGPT, and the Copilot Chat I use all fall into this category. They generally have little hands‑on ability (can't read/process files or generate files); the best they can do is generate images. But that doesn't mean they are useless — you can ask one to check the weather or crawl the entire web for information, for example: "What are the popular conversational AIs on the market, and what are their reputations?"

How to choose: Most people never push AI to its limit. The differences you can feel are mostly in the manufacturer's language style settings (warmer vs. cooler tone), the logic for handling problems (prioritizing reasoning vs. retrieval), and capability boundaries (whether it can generate light documents, context length, companionship features, safety‑restriction differences, etc.). Then there are price and ecosystem. For example, I'm a Microsoft ecosystem user; by paying a bit more I get Copilot, unlocking chat functionality, Tasks capabilities, and Copilot integration in other apps within the ecosystem.

Summary: Chat AI has a low barrier to entry. For most people, there's no huge gap in capability — the biggest difference lies in the conversational experience, unless you are a deep user like me. You should choose based on the style you like (more companion‑like vs. cooler, more talkative vs. less talkative). So if you don't have a strong urge to learn about AI, every platform actually offers some free quota. You can try them out — even use 10 free queries from one platform and 10 free queries from another — hop from one to the other and enjoy the benefits for free.

Tool AI

Form: Specializes in a specific vertical domain, such as video generation, image generation, or code generation. The advantage is that it understands your language in the direction of its own capability, making the information more focused. But its chat functionality is weak — how can a tool chat?

Market positioning: There are many types of this kind of AI. My Copilot embedded in Excel falls into this category. Why is Tool AI the type most likely to replace humans? Because they reduce the "execution barrier" in a vertical domain directly to zero. In the past, you needed to master 80% of Excel functions to complete a complex spreadsheet, or years of artistic skill to storyboard a short drama. Now, you only need clear input. It's not taking away your job — it's taking away the premium that used to go to those "who only mastered basic execution tools but lack structured thinking."

A colleague of mine used to spend 7 out of 8 hours writing code, and still might produce buggy, messy code. Now he spends 1 hour explaining clearly what he wants, then takes it easy. He writes code faster, with fewer bugs. At the same time, AI‑generated short dramas have already reduced the market demand for short‑drama actors.

How to choose: Look at how relevant it is to your work. I work with spreadsheets often, so I don't need image or video generation tools. I use Copilot inside Excel to get most of my work done. Copilot in Excel has cut my spreadsheet‑related work time by at least 70%.

Summary: Choose this type of tool based on your needs — you don't have to have one. But it is the category that truly gives you the immediate feeling that "AI is gradually replacing humans."

Search AI

Form: Usually embedded inside a search engine. It summarizes based on what you search for, giving you more efficient results.

Market positioning: This type of AI often goes unnoticed — because it doesn't converse with you; it just summarizes. You think you haven't given any instruction, but when you type something and hit Enter, it's already activated.

How to choose: There's not much to choose — it depends on which browser you use. You wouldn't switch browsers just for one feature. Moreover, it usually can only summarize the content it finds — if the search results are messy (e.g., ads, misinformation), the summary will be messy too. Don't worry too much about this choice. It's unlikely you would abandon a browser you're comfortable with just because another one has this feature.

Summary: This type of AI is usually free; it's just a feature enhancement that search engines add to their own browsers. It is the only type of AI that "doesn't invite you to participate" — all others involve some interaction with you. Moreover, it is a "non‑reasoning, summarizing AI." To some extent, it can avoid the hallucinations caused by reasoning, which is one reason Perplexity is well‑liked.

Agent AI

Form: This type of AI is usually more than just a dialog box. It connects to your email, calendar, browser, cloud storage, and other tools. You can tell it, "Check tomorrow's meeting schedule for me and send an email reminder to my colleagues." It will then look through the calendar, write the email, and send it — rather than giving you a block of text for you to do it yourself. Some agents can even execute tasks in the background and notify you when they're done. It possesses the closed‑loop ability to autonomously break down tasks and call external tools.

Market positioning: If Chat AI is a colleague who only offers ideas, then Agent AI is an assistant that actually gets work done for you. This is currently the fastest‑growing and most closely watched direction in AI, because it solves the problem of moving from "talking" to "doing." There are already many consumer‑grade agent products on the market. For example, the Copilot Tasks I use, OpenAI's ChatGPT with Agent Mode, Google's Gemini, and Anthropic's Claude have all introduced agents. In China, ByteDance's Coze lets you build your own agents (the previous ones bring the food to your table; this one is more like giving you a kitchen). There are plenty of tutorials online, and the barrier to entry is not too high.

However, this track is still very early. In practice, stability, accuracy, and the ability to handle complex tasks are far from the point where you can "completely let go." I often encounter misunderstandings, operation failures, and situations that need correction. So today's Agent AI is more like an intern that needs supervision, not a fully trustworthy veteran. But even so, it is already a big step ahead of Chat AI — at least it's willing to take action.

How to choose: Choosing an Agent AI is more complex than choosing Chat AI. Its core value isn't "how well it chats," but "what it can connect to and what it can operate." You need to see which of your daily tools it can integrate with. A Microsoft ecosystem user will find Copilot Tasks smooth; a Google ecosystem user will go with Gemini. This isn't grocery shopping — it's buying a phone. You really need to choose carefully based on your habits.

Summary

Agent AI is the most worthwhile AI tool to learn about right now. Tool AI is just a single point; Agent AI is the whole plane (though in vertical domains, it may not offer a better experience than Tool AI). It is the only category that can most fundamentally change "how humans collaborate with AI." It is still very young, and the experience is not yet mature. When choosing, don't look at whose experience is more mature — look at how well it fits your work. Even if you tell me Google's experience is the best, I wouldn't give up the Microsoft ecosystem for it.

Above are the four types of AI tools classified from a user perspective. The fastest to experience is Chat AI; the most intuitive way to feel the changes of the AI era is through Tool AI. But if you are willing to spend time learning and adapting, the efficiency gains from Agent AI are something the other three categories cannot match.

However, Agent AI often has a relatively high barrier to entry, and you may not be clear about its capability boundaries. Jumping in recklessly will only lead to unnecessary expenses. Later, I will write specific practical cases so everyone can directly experience how AI helps with productivity.

Medium：@dskeke

After 11.5M Words of Hands-On AI, Here's What I Learned

熊钊 — Tue, 26 May 2026 15:30:06 +0000

Translated from Chinese.

Preface

I only have a junior college degree, and I work as a lighting product manager — a field completely unrelated to AI. Yet that is precisely where the value lies: if I can do it, so can you.

From March 6, 2026, when I first encountered AI, until now, I have deeply experienced Copilot Chat, with over 10 million Chinese characters of interactive text. I have also deeply experienced Copilot Tasks, with over 1.5 million Chinese characters of interactive text. At the same time, I have conducted extensive interactions on both Gemini and Deepseek. This has given me a very deep hands-on understanding of AI.

Currently, I use AI extensively in my daily life, and it effectively improves my work efficiency. If you are interested in these aspects, you can follow me.

What Is AI? A Machine That Thinks

My conclusion is this: AI is a machine that thinks. You can understand AI as "a person who can think and has extremely broad knowledge." It can turn you into a "beginner" in a field within ten minutes, and a "knowledgeable person" in that industry within an hour. For example, I spent an hour understanding the wedding industry chain: ceremonies, wedding dresses, wedding photos, wedding planning, hotels… which parts are essential needs, and which are "IQ taxes."

If you searched for this content yourself, you would be drowned in the noise of fragmented information across the internet. In contrast, AI can help you integrate and build structured knowledge in a short time. Throw these questions at AI, go back and forth a few times, and you will feel the efficiency of learning with AI.

But we must also be careful: not everything that looks smart is AI. Although many things online claim to be "AI-powered," some are just fixed logic — for example, turning on the heater when it gets cold. That is just a program. AI, on the other hand, does not require you to write rules. You only need to say, "the temperature has changed, you should take corresponding measures." It will think for itself, integrate knowledge, and then tell you whether you should put on clothes or turn on the air conditioner — both are possible.

It can think — that is the real AI. Much of what is called AI on the market today is essentially just automation. Food assembly lines could operate automatically decades ago. Would you call that AI as well?

Will AI Replace My Job? Transform into a "Car Driver" of the New Era

Many people worry that AI will become so powerful in the future that it will replace them. But in fact, history has already presented us with such an era many times — for example, the advent of the steam engine, the automobile, and automation. Society still progressed, and the population continued to grow.

Take the transition from the horse-drawn carriage era as an example. The automobile replaced the "carrying value" of the horse, not the horse itself. Nor did carriage drivers disappear the moment cars appeared. Instead, some of them transitioned into becoming car drivers.

AI will not replace you. But it will be used by those willing to learn to replace "the you who does not learn." A few years from now, if you only complain that "AI took away my job" — what does that have to do with AI? AI has an extremely low learning cost and improves very quickly. There is no need to feel too much pressure. Starting to learn now is not late at all.

Learning AI: How You Express Yourself Matters More

From my experience and journey, I can tell you directly — learning AI has nothing to do with knowledge of programming, math, English, or similar subjects. Using AI well requires more of an ability to express yourself, rather than specific domain knowledge. Over‑relying on deterministic thinking, when facing large language models with emergent and fuzzy properties, becomes a self‑limiting constraint.

As long as you can speak, AI will break down and process your requests on its own. I cannot write code. I only tell it, "I want this effect," and it can achieve it. This may sound a bit mystical right now. AI is not a magical dragon — it cannot fulfill your wish of "give me 1 million dollars." But if you say, "give me a picture of a dog," AI can still do that.

Is Using AI Safe? How to Balance Efficiency and Security

Here we need to discuss how AI works. AI generates content based on: the information you provide + world knowledge + reasoning. If you reveal too much and are overly vigilant at the same time, you will perceive it as dangerous.

You are wearing the uniform of a well‑known local company, speaking the local dialect. If you also casually mention your commuting route and how long it takes, a person with strong reasoning skills could even accurately guess which residential complex you live in. You think they are "watching you," but in fact, all that information was voluntarily provided by you.

As for privacy concerns, that varies by platform. AI is a category, not a single product. Security depends on the platform you choose. Just like cloud storage, social media apps, or even mobile phones — who can be 100% certain they will never be attacked? The main point I want to make is that AI is just one form of software. If you are truly very worried, the best approach is simply not to give AI any important information.

Are AI's Answers Accurate? Understand the Boundary Between Restructuring and Inference

Many people who lack independent thinking treat everything AI says as gospel. In reality, the way (text‑based) AI works can be roughly divided into two types:

Restructuring and summarization — this is the most basic capability. The information here all comes from existing knowledge. AI is merely performing a summary.
Inference and guessing — this is AI's core capability. It makes guesses and inferences about phenomena based on existing knowledge and patterns. But it is only inference, not reality.

Example: I buy a bag of apples. AI thinks about this bag of apples.

Restructuring and summarization: This bag of apples weighs 2 kg. It contains 10 apples. 9 are ripe, and 1 is not yet ripe enough. This is a summarizable reality.
Inference and guessing: These apples are all sweet and taste good. This part is entirely inference and guessing. Because no one has tasted them — even if one apple is sweet, there is no way to guarantee every single apple is sweet.

Regarding control over AI's information, users must have their own standard of judgment. If truly unsure, ask AI to provide the source of the information.

Conclusion: Understand the Car Before the Streets Are Full of Cars

AI is truly a beneficial tool of our time. It is very useful and very quick to learn. In the future, its importance may become as great as the internet's. And right now, AI is still in its early stages. If you want to learn, now is a very good time.

Just like the earlier example of the horse‑drawn carriage and the car. When you see a car, you should already consider learning about it — not wait until the streets are full of cars before you think about acquiring knowledge related to them.

Medium: @dskeke