AI Deep Research — Turning a One-Line Request into a 47-Tablet Comparison in 14 Minutes

I want to buy a tablet.

Not “the iPad is already decided” — I genuinely hadn’t made up my mind. My desktop at home runs a 4K screen, and I spend a lot of time working through remote desktop.

This kind of request — “I roughly know what I want but I can’t articulate it” — is the best possible fit for AI deep research. In the video I run it live in Claude Code: 14 minutes, and out comes a click-to-sort comparison table covering 47 products across every dimension.

Here’s a breakdown of what I actually did.

1. Spell out the starting point until it’s actionable

Step one isn’t putting the AI to work. It’s making clear why I want this thing.

I opened with a short brief: this tablet is mainly for remote-controlling my home desktop (Moonlight / Sunlogin), so —

• Screen has to be 11 inches or larger; 4K over remote needs to look sharp
• Thin and light, high screen-to-body ratio, narrow bezels
• Connectivity has to be solid; built-in 5G is a plus, and without 5G it needs Wi-Fi 7
• Only models released in 2024 or later

Five sentences isn’t complicated, but it’s ten times more specific than “recommend me a tablet.” AI can’t read minds. The granularity of your request determines what it can find.

2. Ask for a usable artifact, not a wall of text

Step two is defining the deliverable.

I said: when you’re done searching, I want a web page. A table with 47 rows for 47 products, one column per comparison dimension — brightness, refresh rate, panel type, screen-to-body ratio, Wi-Fi version. Clicking a column header sorts by that column.

Why be this specific? Because “AI recommendations” have burned me before. Ten paragraphs describing ten products, and I finish more confused than I started. A table plus sorting means I can work it with my own hands until a choice falls out.

Getting the AI to output something usable instead of prose you read once and throw away — that’s the biggest lever there is with these models.

3. Let the AI ask you questions

After I finished talking, Claude Code didn’t immediately run off to search. It asked a few questions first:

• What’s the budget?
• China-market version or overseas?
• Where should the output file go?

My answers: no budget ceiling (let’s look first), any region, output to my notes folder.

This step matters. An AI that never asks anything back is dangerous — it will force a deliverable out of your vague request, and you end up picking from a pile of compromises.

Asking questions means it’s genuinely understanding, not performing understanding.

4. Look at the result, not the process

The research finishes and the 47-tablet comparison page comes out.

I page through it live: OLED versus LCD clearly labeled, refresh rates covered from 90 all the way to 165, thinnest at 4.7 mm, lightest the 11-inch iPad Pro M5 at 444 grams. Click the screen-to-body column and it sorts from 94% down to 83.4%. Every model has a price.

I don’t need to see how the AI searched, how many API calls it made, or which pages it read. All I need is this table.

Wrapping up

AI isn’t about “fast.” It’s that half the work of grinding a request down to the point where you can act on it has been taken over.

The other half will always be on your side.

Source: EP0002_audio.mp3 · ASR model gemini-2.5-pro · full text of the 14-minute original recording

[00:00] Hey everyone, welcome back to Pinpin’s channel. My voice is a little hoarse today — you can probably hear it. So today I’m going to keep sharing how I use AI in my day-to-day work. Today’s share is using AI to do a detailed piece of research. Let’s get started right away, because research work does take a fair amount of time, so I need to leave time for it to run. And I’m doing this research through Claude Code, in that terminal form.

[00:30] Oh, I — so, when I’m talking to Claude Code I always use a voice input method. It’s an input method I built myself. If anyone’s interested, leave a comment and I can share it with you to try out and use. I’ve been out of the house a lot lately, so I use a pad to remote into my computer at home and get some work done with AI. So I’ve been thinking about buying a new large-screen tablet — and that’s what made me think of having AI do some research.

[01:01] So we start the natural way: first I have to tell the AI clearly where I’m coming from, what my goal is, so it can do the job better. Like: I want to buy a tablet, and the main use is remote control — Moonlight or Sunlogin, that kind of thing — to control my computer at home. So this tablet should ideally be 11 inches or bigger, very thin and light, ideally with a high screen-to-body ratio, meaning narrower bezels, which looks better. As light as possible, and thin too. And I want products from 2024 onward, and ideally a sharper screen too, because my computer at home runs at 4K resolution, so when I’m remoting in it’ll look clearer. And since this thing is mostly for remote control, the networking should be good — built-in 5G is a plus, and if there’s no 5G, then ideally a Wi-Fi 7 version.

[02:10] Then I need you to build a multi-agent team, do deep search and research on the internet, and assess how credible the sources are — don’t just grab whatever information comes along. And I need you to report back to me as a web page. On that page there’ll be one big table, where every row is a tablet and every column is a spec you can compare across these tablets. The things I just mentioned obviously have to be in there. And then the details too, like its length and height in millimeters, its brightness, its refresh rate, whether the panel is OLED or LCD, all of it has to go in. So every comparable spec you can think of goes in there. And the table should be sortable by clicking — when I click a column header, say I click screen-to-body ratio, it sorts by screen-to-body ratio, which makes it easier for me to pick between these pads. I don’t know if you got all that — if the requirements aren’t clear, you can ask me questions and I’ll clarify.

[03:30] So you can see I’ve already sent that requirement in, it went out automatically. Claude Code will now analyze the requirement, and it may come back to me with some questions. I don’t know how everybody else usually uses AI, but the nice thing about a local terminal agent like Claude Code is that it really can call search and intelligently assess where the information is coming from, and it can also present the research results much more clearly as a web page. That’s different from what we normally use, like Doubao or DeepSeek — those are just large language models. And we can see, OK, here are the questions I didn’t cover: it’s asking me what my budget is, and I won’t set an upper limit for now.

[04:30] OS — OK, I’ll pick no restriction, I’ll decide after the comparison. Region version — mainland models and overseas models are both fine. And delivery: I like having it output into a folder I use a lot, so let’s have it drop this into my notes folder.

[04:51] You can all see it understands our requirement, does some analysis, and after checking back with us and confirming the requirement it got is accurate, only then does it start executing. And its output, like I mentioned, comes out as a web page — it isn’t just text, so the result it produces is much clearer, and it’s easy for us to share with other people.

[05:32] OK, the results are back now, let’s take a look at what the research came up with. On this page you can see a deep comparison of 11-inch-and-up tablets released from 2024 through 2026, and it fits my scenario of remote-controlling a 4K desktop. It built a weighted average score — screen, networking, portability, performance and price. Let’s skip the rankings that came out of that for now, and look at the very detailed head-to-head table of 47 products. You can quickly filter by operating system, filter by screen, whether the screen is OLED or LCD, whether it supports 5G, and whether it has Wi-Fi 7 — all of it filterable right here.

[06:17] And you can see the specs in here are very detailed: screen size, resolution, PPI, aspect ratio, refresh rate, right — we can see refresh rates from 90 all the way to 165 — then the screen’s peak brightness, touch sampling rate, then its length and width, screen-to-body ratio, right. Screen-to-body goes from 94% all the way down to 83.4%. Tablet thickness — let’s look, at the far end there are 8.5, 8.5 mm ones, and the thinnest is 4.7 mm, right. On weight, if I want the lightest, that’s 444 grams, the 11-inch iPad Pro M5, and that’s the lightest one in here. Plus the AnTuTu benchmark scores in here, then the storage situation, battery capacity, right, its charging wattage — all of it visible right here. So it makes it really easy for me to do a — including its prices, right — really easy for me to filter and choose.

[07:23] So that’s it for today’s share. If you want to know more about using AI, follow me, and we’ll post something every day.