An AI Checkup for Your PC — 30+ Windows Bottlenecks Scanned in 5 Minutes

A few friends asked me recently whether sharing a new Claude Code use case every day means I’ll run dry fast.

I won’t. By my estimate the tricks I already have are good for six months.

Today’s topic is a widespread misconception — most people treat Claude Code as a coding tool. Yes, it can code, but what it’s genuinely strong at is using programming to accomplish all kinds of things. The gap between those two framings is a hundredfold.

This episode, I put it to work as my computer doctor.

1. Why not use a PC manager utility

I have a Windows 11 LTSC workstation at home, and I’ve been through every flavor of “XX Guard” and “XX Manager.” Two problems:

• They eat system resources — they’re sitting in the background running themselves
• They’re black boxes — you have no idea what they’re actually doing

Claude Code is different. Every step is visible, and it can tune deeply for your specific machine — not a generic script written for ten thousand people.

2. Step one: get the starting point aligned

I didn’t have it start working immediately. First I laid out a brief:

“Optimize this computer’s performance for my work habits and work scenarios. Start with a full audit of the configuration and installed software, do a complete review combined with what you know about me, and then decide how to optimize.”

One of the most important things about using AI is giving it the goal up front. That way it keeps recapping that goal throughout, and it drives what knowledge it pulls and what it searches for.

An AI that never asks anything back is dangerous. One that asks is actually trying to understand you.

3. A contrarian finding: more RAM allocated to VRAM isn’t better

It ran a full diagnostic first — CPU, RAM, VRAM, startup items, redundant software, all swept.

One thing I spotted: I’d allocated too much system RAM to VRAM — could that actually be hurting performance? I edit video in CapCut, so I’d assumed giving VRAM a bigger allocation was the better configuration.

Its explanation surprised me: modern Windows applications borrow system memory for the GPU automatically. Manual allocation like mine only helps pre-2018 games, older CUDA programs, and some Adobe plugins. Apps like CapCut peak at only 4-8 GB of VRAM anyway.

“You’re better off leaving most of that memory as memory.”

Newer drivers aren’t necessarily more stable either. It tested my machine and found the older driver was actually more stable.

4. Multi-agent concurrent research on the last 3 months of optimization findings

Once I understood that, I had it spin up a multi-agent team to search the web for verifiable optimization best practices from the previous three months.

The research came back as a checklist:

• Revised VRAM allocation
• One driver rollback recommendation
• BIOS update (with steps specific to my machine model)
• Startup item trimming
• Uninstall IE
• Anti-ransomware ASR configuration
• Laptop power management plan

Deeply researched, highly specific. No ordinary PC manager utility gets to this level.

5. Execution split across 3 conversations + an HTML to-do list

It understands my working rhythm well, and suggested splitting into three conversations:

1. T0 — everything PowerShell can do directly, no restart required
2. BIOS update — needs several restarts, so run it separately
3. In-app software settings — can be slotted in whenever, not urgent

Then I had it write everything beyond T0 into an HTML document — each recommendation carrying its reason / performance benefit / risk / firmware download link.

Open the page and the whole optimization checklist is right there. That same file also works as context for the next conversation.

Wrapping up

Once you do this once, it knows your machine. After that you can save it as a scheduled task that runs in the background periodically, with nobody watching over it.

That’s what Claude Code can do — beyond coding. It does it through programming, but what it accomplishes isn’t programming.

Harsher than 360, more transparent than a PC manager.

Source: EP0007_audio.mp3 · ASR model gemini-2.5-pro · full text of the 8-minute-59-second original recording

[00:00] Hey, I’m back again. A lot of friends have asked me these past couple of days: if you share one way of using Claude Code every day, are you going to run out of material, are there really that many ways to use it? But I’m honestly not worried at all. I did the math, and with all the different things I’ve done before, I figure sharing for half a year is no problem. So what I’ve got for you today also comes out of a common misconception I’ve noticed: most people treat Claude Code as a programming tool. Yes, it can program, but it uses programming to get all kinds of other things done.

[00:30] So what I want to share today is using Claude Code to maintain and manage my computer. In the past I’d use all kinds of software — whether it’s this X Guard or that X Manager — and those tools all eat a certain amount of system resources, and frankly we don’t really know what they’re doing behind the scenes in the system. But once I had Claude Code, I found that a lot of the cleanup and maintenance work on a computer is actually something it can do better. So let me just demo it for you. Still my very favorite voice input method, let’s take a look.

[01:04] Where I’m coming from is that I want to optimize my computer’s performance for my work habits and work scenarios. So I want you to first do a full check of my computer’s configuration and what software is installed and being used, and combine that with what you know about me to do a complete review. And after the review, we’ll look at how to optimize the computer’s performance settings together.

[01:31] OK, so I gave it a request like that first. A really important part of using AI is syncing where you’re coming from, your purpose in doing this, over to Claude Code first. That way it keeps recapping that goal through the whole process that follows, so it can pull up the matching body of knowledge and run the relevant searches to hit that core goal. And in the process that follows, we’ll speed up its thinking a bit, but I’ll also show you the whole process without cutting a single frame.

[02:03] OK, the data collection is done now. It’s already seen my machine’s full configuration, including my RAM and VRAM settings, plus startup items and some redundancy in my software installs. In a situation like this, generally speaking, here’s what I’d do next — right, from a rigorous computer-performance-optimization standpoint, thinking about it for a second, there are probably a few angles. The first is clearing out junk files. Then clearing out software that sits in memory but doesn’t get used much, or some of the software running right now, right. And then we can also optimize some of the computer’s system settings, especially since mine is a Windows 11 LTSC build.

[02:44] So for these things, for my computer’s current state, I want you to spin up a multi-agent team and search the whole web along the dimensions I just mentioned, but not limited to those, and look at experience from the last three months, or best practices, that clearly improve a computer’s performance. Then based on those best practices we’ll decide which optimizations to actually do, and I’ll make the call. And in the middle of this, I actually spotted an issue — I think I’ve allocated too much VRAM to RAM, too much RAM to VRAM, and that might hurt performance. So I also want you to explain the reason for that in some detail.

[03:28] See, I normally use CapCut to make videos — so is allocating 32G of RAM to VRAM a more sensible setup? Or can CapCut not even call on that much VRAM to support making high-resolution video? I need you to give me an explanation in plain language.

[03:50] OK, see, based on its feedback plus some of my own thinking, we hand that result back to it, and then based on the whole computer configuration it just checked, including my thinking and suggestions, it’ll do a concurrent search and research pass. It’ll use the latest verifiable community experience to help me optimize my computer configuration and tune performance. And we’ll speed this stretch up again.

[04:18] OK, it gives me an explanation first — it didn’t start the research right away. The explanation it gives me is that allocating RAM to VRAM is actually the wrong approach. It explained that these days applications on Windows will lend RAM to the GPU for display use anyway. Doing it my way only helps in certain specific cases — like, what I can think of, and what it just mentioned, are some pre-2018 old games, plus some Adobe plugins and old CUDA programs that use large VRAM.

[04:49] There might also be some AI applications. But apart from those, in common use cases, including CapCut, peak VRAM usage is probably only in the 4G to 8G range. So what it actually recommends is leaving most of the RAM as RAM.

[05:07] OK, I understand the explanation you gave. So now spin up the multi-agent team and search concurrently for those optimization findings I mentioned.

[05:20] OK, the research is over, it’s back. We can see it’s already laid out a checklist. The first — let’s go through them one by one, and it really is very professional. The first is adjusting that VRAM setup. Then the second, it says one of my drivers, the newer driver is actually worse than the older one, so it recommends I roll that driver version back. And then BIOS, it recommends I upgrade. That’s already a very deep recommendation, a very targeted and very deep recommendation. Most computer-management software wouldn’t go that far.

[05:45] Of course that’s also a relatively advanced operation. If you really want to go do it, you’d still need Claude Code to give you further detailed guidance, if you’ve never done a BIOS upgrade before. Including some of my startup items, and the boundaries of some software optimizations, and the Windows build can be upgraded, right, and it laid out the corresponding payoff too. And at the same time, you can see it’s uninstalling IE. Fine, uninstall IE. Then let me see what else there is. Some anti-ransomware ASR recommendations, plus a laptop power management plan.

[06:23] OK, see, it knows my working rhythm really well. It says that given my agile work rhythm, it recommends splitting this task across three conversations. And yeah, it’s right — look, the first conversation is the stuff it can do right now with a PowerShell script, so no reboot needed, right. Then the second part is about the BIOS upgrade, which means rebooting several times. And the third is settings inside applications, right, which can be configured. So the third can be slotted in whenever, no rush. OK, let’s do it that way. So I’ll give it a piece of feedback.

[06:48] So in the current session, finish all the T0 work and the T4 part you recommended, and for the T1 and T2 parts we haven’t done yet, please write a detailed HTML web document, so I can see very clearly the optimization to-do list you recommend, the reasons for the recommendations, and their performance payoff and possible risks. So we do a checklist-style review, which can also be left for the next conversation as a reference. Sound good? OK, let’s do the current session’s work first.

[07:26] OK, you can see it’s already finished. It tells me the T0 items it ran automatically are done, and then there are two things it needs me to do by hand — run this Windows update, and then this Defender script, which needs me to execute manually. And then there are two pieces of software, one of them is installed twice and one of the two versions can be uninstalled. Oh, and at the same time it says the HTML file is already open, so let’s take a look.

[07:52] OK, we can see it’s opened a web page now. It’s a more visual way of telling me what it’s already done. And then there are a few operations that need to be done by hand, and for the VRAM configuration, including the Windows update and the BIOS update, it laid out a detailed procedure — including the payoff and the risk, and how to do it, and it’s already given me the firmware download link.

[08:22] OK, I won’t go through it in detail. So you can see a process like this was completely tailored to my computer’s configuration, finding the latest drivers and optimization approaches and setting them up. And of course there’s the everyday stuff too, like clearing out junk files — it can absolutely do that. And once we’ve done this once, it actually saves the experience, and after that you can turn it into a scheduled task, meaning every so often it can automatically do this in the background, with nobody watching over it.

[08:49] So that’s what Claude Code can do — beyond programming, beyond programming. It’s using programming to do it, but what it delivers is a set of very useful features that aren’t programming.