If You Can Manage People, You Can Use AI—A Real Class-Prep Run with Claude Code as a Digital Employee

Here’s what I’ve been noticing lately: everyone says to treat AI as a digital employee, but when people sit down with AI they still use it like a search engine—a few keywords, or one bare question, never talking to it like a person.

So today I’ll walk through how I put AI to work as my employee on a real job.

If you had an assistant, how would you brief them?

I was recently invited to give a full-day, 6-hour workshop in person. You’re the instructor, the brief just came in, and you have an assistant. What do you have them do? The normal move is to get the organizer on the phone: who’s in the audience? What problems are they running into? What value do they want, and which cases and methods are they leaning toward? Where’s their current level? And if there’s a pre-class survey, interviews, or attendee background material, can we get it up front and go through it together?

Then the material arrives and there’s a lot of it. No instructor can digest that fast alone, so you’d ask the assistant to boil it down and hand it back as a readable report.

Now take that exact thinking to Claude Code and have it do the assistant’s job. The pre-class survey spreadsheet the organizer exported, plus the Get Notes transcript from the online session where we sounded the attendees out—dump all of it into the chat box.

Briefing the job in three sentences

Before making the request, stop and think: if I got one shot at briefing this assistant, how would I put it? Here’s how I usually do it—

I’ve been invited to give a full-day, six-hour workshop in person. What I’m giving you here is the attendee material, the speech-to-text transcript from the needs research we ran with the attendees beforehand, and the survey form I had them fill out. Read through all of it first and organize it. To prepare the course content better and give the attendees takeaways they can use, I need you to produce a pre-class audience needs analysis report as a web page in the light PPVI style, so I can see who the audience is, what they need, and what level they’re at. Then analyze the challenges I may hit while preparing the course content, and give me an initial proposal for solving each of them.

Break that down and it’s three things: what material I’m giving it; what the purpose and motivation of the work is; and in what form and to what standard it should deliver. That’s the three-part request framework, stripped to the bone.

Where agents and models diverge is everything after this step. A good agent with a good harness configuration will find and inject the skills and prior experience relevant to the job on its own—things I never told it. A good model brings stronger reasoning and expression, longer context, and a deeper base of existing knowledge. It plans carefully before it starts work, and it reaches for a richer set of tools: researching the organizer on the side, even calling an image model to add illustrations.

Mid-run I added one more instruction. I was recording, so any private information on the page, names included, had to be redacted. And once the page was done, generate a one-page PDF I could hand straight to the organizer—that part used onepage-pdf, a skill I open-sourced earlier.

What it handed back is no longer assistant work

Here’s the result: a web page plus a PDF. The page looks great. It leads with three core conclusions, then gives a clear audience profile—concentrated demographics, scattered industries, layered skill levels. They’re mostly women born in the 1990s living in Shanghai, but they come from a dozen-plus industries. Some are still at the ask-a-question stage; others already have agents they orchestrated themselves running inside their business.

The pain points cluster tightly, and they’re exactly what I keep saying on this channel: AI has to serve real business and get built into real workflows. Right now they’re mostly on Chinese models and tools. The report pulled the interviews and surveys into six core needs. One is drawing the boundary first—how far can AI actually go? Another is where Chinese models and tools fall short of the strongest agents and models like Fable 5. It also profiled representative attendees, gave me the prep challenges and matching plays I’d asked for, and closed by laying out an action list for me.

That’s not assistant work anymore. That’s what a partner does—it was even assigning work to me.

I couldn’t resist a side-by-side: same model, different tool

While it was generating the page, I couldn’t keep my hands still and did one more thing: wired Fable 5 into Tencent WorkBuddy. Same request, same material, same model—what difference does the tool make?

I regretted it the moment it finished. Forget output quality; one web page burned through tens of yuan in tokens.

In fairness, WorkBuddy produced its page faster (we didn’t ask it for a PDF), and the content covered what it needed to and was genuinely well done. The page design was competent if unremarkable. It used clear charts the same way, distilled six pain points, listed the hottest audience questions, and drew out implications for course design. What it couldn’t do, thanks to tool-calling limits, was layer in anything search and reasoning would have added—and the visuals weren’t as striking.

Compute costs will eventually exceed people costs

Plenty of Chinese companies are already rolling WorkBuddy out across their whole workforce to make people more productive, and more than a few give employees big compute allowances wired to top-tier models. Judging from today’s test, even on a model like Sonnet rather than Fable 5, delivering a document like this costs at least a dozen-plus yuan in compute. That’s steep.

So somebody inside the company should own this: deciding which work is worth backing with compute that expensive, and how much business growth it returns.

One day, a company’s compute costs will exceed its people costs. I believe that day is no longer far off.

Source: EP0057_audio.mp3 · ASR model gemini-2.5-pro (chunked parallel) · full text of the original recording

[00:00] I’ve had an insight lately. Everyone says you should treat AI like a digital employee, but most people, when they actually sit down with AI, still use it like a search engine—a few keywords, or a bare question, fired straight at it. Prompted that way, AI doesn’t talk like a human at all. So today I want to share how, in a real piece of work, I use AI as my employee. I got invited recently to teach a big full-day class in person. As the instructor, when you get the brief, if you had an assistant, how would you hand out the work? The normal line of thinking usually goes like this. You get in touch with the organizer of the course

[00:26] and you ask: who’s in the audience this time? What problems are they running into? What value do they want to walk away with? The cases and the methods they want to hear—which way do those lean? Roughly what level are they at right now? If there’s a pre-class survey or interviews, and any materials on the students, can we get those first so we can analyze them together? And then once you have all that material—once you have all that material, the volume of it can be huge. In a short window, there’s no way the instructor can get through it fast enough on their own to get

[00:51] So we’d ask the assistant to digest the material down to the gist and hand it back to us as a report we can take in at a glance. Now let’s take exactly that line of thinking and talk to Claude Code with it—let it do that assistant’s job, that employee’s job. First, the pre-class survey: the organizer exported it into a spreadsheet and sent it over. You can see there’s a ton of content here, and it’s a real pain to read. On top of that, there’s the online meeting where the students talked through what they wanted from the course. That discussion is a Get Notes transcript, and I need to copy all of it

[01:17] and drop it into this chat box. I can even dump everything the organizer gave us in here, all at once, without even thinking about it. Before I make the request, let’s stop and think: if I only got one shot at talking to this assistant, one chance to hand over materials, how would I phrase the ask? Here’s how I usually do it. I’ve been invited to give a six-hour, full-day class in person. What I’m giving you is the students’ materials: the pre-class needs research on the students,

[01:42] the speech-to-text transcript of it, plus the survey form the students filled in. Please read through all of this and organize it first. To prep the course content better, so the students walk out with something they can actually use, I need you to give me a pre-class student needs analysis report as a web page in the PPVI light style, so I can get a better read on who the audience is, what they need, and what level they’re at right now. And while you’re at it, analyze

[02:07] the challenges I might hit while preparing this course, and for each of those challenges, give me a first-pass proposal. That’s it. So you can see what I told it. I told it exactly what the materials I’m handing over are. I told it the goal, the starting point of the work. And I told it the final format and the standard for delivery. Those three things are the bare bones of how you frame an ask. The differences between agents and models actually show up after that step. A good agent, a well-configured

[02:32] harness, will automatically find and inject the skills and the experience relevant to this job—the stuff I never told it. And a good model reasons better, says things better, holds a longer context, and can draw on its own richer built-in knowledge to make a tight plan before the work even starts, plus it can call on a wider set of tools. It can go do extra research, for instance—it can go research who the organizer is—it can even call an image model

[02:58] to add some illustrations. Oh, I actually forgot something. We’re recording a video right now, so the content has to be redacted. So I’m firing off one more instruction. Sorry, I forgot to mention this: this is all going out as a video, so in the web page you hand me, anywhere there’s a name or other private information, it needs to be redacted. Once the page is done, generate a one-page PDF so I can hand it straight to the organizer as a reference for them. You can see it’s already started thinking, and now I’ve thrown one more instruction in

[03:23] This also uses that skill I open-sourced a while back called onepage-pdf. You can go watch the videos on that to learn it, or search Zhang Pinpin on Google, find my blog, and download the transcript of this video. Let’s fast-forward now and look at the result. Okay, it’s finished, and it gave me a web page plus a PDF. Let’s open the PDF. The presentation is honestly pretty good-looking. It leads with the conclusions—three core ones—then a very clear student profile. The group is concentrated, but the industries are scattered and the skill levels are

[03:48] tiered. They’re mostly women born in the 90s living in Shanghai, but they come from a dozen-plus different industries. Some are still at the ask-a-question stage; others already have self-orchestrated agents running inside their business. The pain points are very concentrated, and they’re exactly what I keep hammering on this channel: AI has to serve real business, it has to get built into the actual workflow. What they use today is still mostly domestic models and tools. And it pulled the interviews and the survey together into six core needs. For example: first, get clear on the limits—

[04:14] how far can AI actually go? And where exactly do domestic models and tools fall short of the strongest agents and models, of something like Fable 5? Then it analyzed typical student cases, and as I’d asked, it laid out the challenges I might hit while prepping plus a proposed play for each one, and gave me an action checklist. This isn’t an assistant’s work anymore—this is what a partner does. It’s even assigning work to me. While I was waiting for that page, I couldn’t help myself and did something dumb: I hooked Fable 5 into WorkBuddy.

[04:39] I wanted to know: same brief, same materials, same model—how much difference does the tool make? I regretted it the second the run finished. Never mind the quality of the output—just producing one web page burned a good few dozen yuan of tokens. WorkBuddy actually generated its page faster, because we didn’t ask it for a PDF. And in fairness the content is good enough, done pretty well. It’s just that with the limits on tool calls, it couldn’t fold in much of the extra information you get out of search and reasoning. And it couldn’t pull off a particularly

[05:05] beautiful visual presentation. But I think it’s good enough. Let’s take a look. The page design is solid, middle of the road, and it uses clear charts too. The overall content structure also distills six pain points, lists the hot audience questions, and gives me takeaways for course design. A lot of domestic companies already lean heavily on WorkBuddy to get their staff up to speed with AI, and some of them do give people a big compute allowance and hook up top-tier models. But going by today’s test,

[05:30] even if what we hook up isn’t Fable 5 but something like Sonnet, delivering a document like this costs at least ten-plus yuan of compute. That’s still quite high. So inside a company’s actual business: what kind of work is worth backing with compute that expensive, and how much business growth does it really bring? That’s genuinely worth putting someone on full-time to think through and design. Because one day, a company’s compute cost will exceed its headcount cost. And I believe that day—I believe that day isn’t far off at all. That’s it for today.

[05:55] See you tomorrow. Bye.