Claude Code Might Be the Worst-Named Product in AI, and a Concerning Number of Smart People Think It's for Developers.
I Thought Claude Code Was for Developers. The Productivity Leap Is Bigger Than ChatGPT Ever Was.
A few years ago I had a lot of opinions about ChatGPT. Which was impressive, because I had barely used it.
I knew some of the people building it. I had employees using it every day. I’d read the articles, listened to the podcasts, developed nuanced takes. I could explain the risks, the opportunities, and the future of AI with the confidence of someone who absolutely knew what they were talking about.
I was reviewing a wine I’d only smelled.
Then I finally sat down and used the thing. Within a month I was completely dependent on it. It had quietly rewired how I work. The gap between what I thought it was and what it actually was turned out to be embarrassingly large. I wasn’t a little wrong. I was the kind of wrong that should come with a refund.
So when Claude Code showed up and most people immediately filed it under “developer tool, not for me,” I recognized the exact shape of my old mistake. It was the same confident opinion about something I hadn’t actually tried. For once, I skipped the opinion and went straight to the experiment. Thank God I did.
Yes, it was built for coding. That part was true. And yes, I’ve mostly used it for coding-related work. I’ve built most of Oomira inside Claude Code without writing a single line of code myself, which is either an incredible endorsement of AI or a deeply concerning review of my technical abilities.
But what became obvious almost immediately was that “coding tool” is a wildly misleading description of what this thing actually is. Calling Claude Code a coding tool is a bit like calling Netflix a DVD company. Technically accurate, at least at one point in time. Completely misses the point.
What Claude Code really gives you is a dramatically better way to get a computer to do things. Not code things. Things. And once you see that, the web chat versions start to feel a little bit like explaining your problem to a very smart person through a mail slot.
It’s a Surface, Not a Coding Tool
What Claude Code really is, underneath the name, is a much better surface for getting a computer to do things. Not coding things. Things.
The mistake I made was hearing the word “code” and assuming I understood what category it belonged in. It’s the same mistake I made with ChatGPT. I saw the label, filed it away mentally, and moved on.
I smelled the wine and decided I knew the vintage.
What Claude Code actually does is remove a bunch of weird little barriers we’ve all accepted as normal. Imagine never uploading a file again. No dragging PDFs into boxes. No exporting spreadsheets. No digging through your Downloads folder trying to remember whether the file you’re looking for is called “Final,” “Final Final,” or “Final Final Investor Version 3.” The files are already on your computer, and Claude can work with them where they already live.
That sounds boring until you use it. Then going back feels like printing an email so you can fax it to yourself.
And that’s just the beginning. Imagine asking a question about your Stripe data and not just getting an answer, but getting a dashboard built around the answer. Imagine giving it an investor diligence request and saying, “here’s every file I have, build me the folder I’d send back,” and then watching one appear. Imagine asking it to review a thousand documents without hearing the phrase “you’ve exceeded your upload limit.” Imagine turning your customer pipeline into a portal for your team. Imagine describing a workflow in plain, clumsy English and having working software emerge from the other side.
That’s what’s hiding behind the label “developer tool.” Not a new career. Not a personality transplant. Not a sudden urge to start explaining Kubernetes at parties. Just a dramatically better way to interact with a computer. And it speaks English.
The Next Floor Is Always “For Someone Else”
The funny thing is that every floor on this staircase initially looked like somebody else’s job.
A few years ago I would’ve told you I understood AI. Not deeply, but enough. I knew what ChatGPT was. I knew roughly how it worked. I had opinions. Caveats. Concerns. Predictions. Then I actually used it and discovered I was standing on the first step of the staircase giving tours of the building.
That’s the trap. The first level is so impressive that most people assume they’ve seen the whole thing. I did. Then I found another floor. Then another. Then another. Every single time, the next floor looked like it belonged to somebody else.
Level 1: Answers. Claude. ChatGPT. You ask questions, it answers. Research, drafts, explanations, summaries. This is where almost everyone starts, and where most people stop.
Level 2: Delegation. Agents. Cowork. Task runners. Instead of feeding prompts one at a time, you hand over an entire job and come back later. You stop driving and start assigning.
Level 3: Action. Claude Code. Codex. Your computer starts doing things instead of talking about doing things. It can access files, use applications, connect systems, build things. This is the level most people immediately categorize as “for developers,” which is exactly why I’m talking about it, because I almost skipped it for the same reason.
Level 4: Publishing This is the first time something leaves your computer. You put a dashboard on the internet. A portal. A tool. A workflow. Something other people can actually use. This sounds like a small distinction until it happens to you. It isn’t. The jump from “my computer can do things” to “I can create software” is one of the biggest mindset shifts on the entire staircase.
Level 5: Integration. The API. You stop using AI as a destination and start wiring it into your own workflows. Your own tools. Your own systems. Your own business.
Level 6: Infrastructure. MCPs. Skills. Tools. You stop building with the blocks and start building the blocks. Other people install the things you make. You’re officially load-bearing. This is where I’m currently spending most of my time.
Level 7: Intelligence. Training. Fine-tuning. Actually shaping the models themselves. The engine room. And before you immediately decide this one isn’t for you, notice that you’ve now heard me say that about every floor above the one you’re currently standing on. That’s the pattern.
The point isn’t which level you’re on. The point is recognizing that the next level almost always looks like somebody else’s job. That’s why most people never reach it.
ChatGPT looked like it was for AI people. Claude Code looked like it was for developers. Putting software on the internet looked like it was for founders who could code. Training models currently looks like it’s for researchers. I’ve been wrong about every floor so far.
The people quietly creating absurd productivity gaps aren’t necessarily more technical. They’re just less willing to accept the label on the door. They see something that appears to be for somebody else and ask a different question: what if it’s for me? More often than you’d think, the answer is yes.
Everyone Talks About Context. Almost Nobody Knows What It Means
Notice what none of those levels measure: whether you actually understand how any of this works. That’s intentional. The staircase measures what you’ve done, not what you know, and those two things are much less connected than most people think.
You can get surprisingly far down this staircase with only a vague understanding of what’s happening under the hood. I know because I did exactly that. A year into building a company with these tools, I still couldn’t have given you a clean explanation of tokens, context windows, model training, fine-tuning, retrieval, or half the other words people casually throw around on LinkedIn as if everyone attended the same secret meeting.
And honestly, that’s fine. Understanding is not an entry requirement. The staircase never once checks your ID.
But don’t take that as permission to spend the next three years waving your hands around and calling everything “AI magic.” The people who understand what’s happening under the hood generally move faster. They make better decisions. They know which tool to reach for. They know why something worked, why it failed, and what to try next. Most importantly, they have a much bigger map.
If you’re spending hours a day with these tools, it’s worth learning the basics. Not a PhD. Not a computer science degree. Just enough to know what a token is, what a context window is, the difference between prompting, retrieval, fine-tuning, training, APIs, agents, and MCPs. Enough that when somebody says those words you don’t nod thoughtfully while secretly hoping nobody asks a follow-up question.
Dharmesh Shah has the best starting point I’ve found for this. Not because it’s advanced. Because it’s the opposite. It starts from zero and explains the fundamentals in plain English, which is a much rarer skill than most people realize. Dharmesh’s beginner AI video.
Do first. Then understand. Then go back and realize how much more was possible than you thought. The first time I understood what a context window actually was, a bunch of things I’d previously blamed on “AI being weird” suddenly made perfect sense. The same thing keeps happening every time I learn another piece of the puzzle.
The One You’ve Already Decided Isn’t For You (I’ve Been Avoiding It Too)
Before anyone panics: no, I am not suggesting every founder needs to disappear into the woods and emerge six months later having trained a frontier model. Nobody is asking you to raise a baby AI in the garage.
What I am suggesting is that you stop deciding things aren’t for you before you’ve tried the smallest possible version of them. That’s the mistake I’ve been describing this entire article. ChatGPT was “for AI people.” Claude Code was “for developers.” Model training is currently “for researchers.” Maybe. But I’ve been wrong enough times that I’m becoming reluctant to make that call from the outside.
The reason to keep moving isn’t because everyone should become an AI researcher. It’s because every step expands your range. When I was sitting at Level 1, there were entire categories of things I didn’t know were possible. Level 3 changed what I thought a computer could do. The first time I put something on the internet, it changed again. Every floor expands the map.
You don’t stop using the earlier levels. I still use all of them, constantly. What changes is that you become a more intelligent consumer of the whole stack. You know which tool belongs where. You know what’s easy. You know what’s hard. You know what’s possible. And perhaps most importantly, you stop confusing the edge of your experience with the edge of reality.
So if you want the smallest possible taste of Level 7, don’t read three books about model training. Don’t listen to five podcasts. Don’t watch twenty hours of YouTube. Go fine-tune a model. Seriously. Spend an afternoon doing it badly. Train something tiny. Teach a model your writing style. Give it a narrow task. Break it. See what happens.
The point isn’t becoming an expert. The point is stopping at the wine tasting instead of writing a review from the parking lot. Because that’s been the recurring mistake of this entire story. The next level always looks like somebody else’s job, right up until the moment you try it.
Tonight’s Homework (The Only Homework)
Install Claude Code. That’s it. That’s the homework.
Open Terminal. Yes, you have one. On a Mac, hit Command-Space, type “Terminal,” and hit Enter. It’s the spooky black window you see in movies. It’ll probably open white the first time, so go into settings, make it black, and enjoy your brief transition into someone who looks like they know what they’re doing. It’s a text box that does what you tell it.
Paste this and hit Enter:
curl -fsSL https://claude.ai/install.sh | bash
Now here’s the honest part nobody tells you. There is a very good chance that line does not just work. Your computer will likely fire back a wall of red text and forty-seven reasons it cannot possibly continue. It needs Xcode. It needs something called Homebrew. It needs Node. It needs a PATH. It needs a blood sacrifice to a god you’ve never heard of.
This is normal. It’s just what happens when a computer that has never done developer things meets a person who has never asked it to. The machine is confused. You’re confused. Everyone is doing their best.
And this is where most people stop. Not because they’ve hit a technical wall, but because they’ve encountered the first piece of evidence that they’re standing in a room they previously believed belonged to someone else. Developers. Engineers. Technical people. People who know what a PATH is. The whole point of this article is that this assumption has been remarkably unreliable.
So here’s the trick. Screenshot the error, paste it into Claude, do exactly what it tells you, and run the command again. Hit the next wall, screenshot, paste, repeat. You are not debugging. You are forwarding your mail to someone who reads it for you. Fifteen minutes of that and you’ll probably be in. I say this as someone who nearly quit on the first error message and has since repeated this process enough times that I should probably be embarrassed.
Once you’re in, type claude and log in when the browser opens.
Then point it at something real. Not a tutorial. Not a toy project. Not a website for a fictional bakery. Pick the thing that’s been rotting on your list for months. Build a website. Analyze your sales funnel. Review a data room. Create an investor dashboard. Build a portal for your team. Automate the thing everyone complains about every Monday. I genuinely do not care what it is. Just make it real. Describe it in plain, clumsy English and see what happens. Because that’s the moment this stops being a clever technology demo and starts becoming a different way of working.
And if you run a company and you’re not using this yet, let me be unusually direct. You have the most to gain and the least excuse. The founders I know who’ve embraced these tools are quietly operating with a completely different range than they had a year ago. They’re building things instead of talking about building things. Answering their own questions instead of waiting for answers. Testing ideas instead of scheduling meetings about whether they should test ideas. They’re not necessarily smarter, and they haven’t become engineers. They’ve simply stopped assuming that the next level belongs to somebody else.
You are not behind because you can’t do this. You are behind because you haven’t tried.
Nobody’s Done. Me Included.
I just spent several thousand words laying out a seven-level staircase. I am absolutely not at the bottom. In fact, I’d bet money the staircase isn’t even the whole shape. Every time I think I’ve got a handle on this stuff, I discover another room, another floor, or another thing I’ve been confidently ignoring because it sounded like somebody else’s job.
So the prescription is the same for all of us, me included: stay curious past the point where you feel competent. Competence is where most people stop. Take one more step than feels necessary. That’s the whole game.
And one slightly disloyal tip before I go. If you and ChatGPT broke up a while ago and went all-in on Claude, it might be time to start seeing OpenAI again. I say this as someone who hits Claude Code’s usage limits and then sits there staring blankly at my screen, wondering what exactly I’m supposed to do with the rest of my afternoon. So I installed Codex. It’s impressively good. Better at plenty of things. And honestly, it was a relief to realize I didn’t have to build my entire future around a single ecosystem and whatever mood it happened to be in that week. Anthropic, somewhat accidentally, solved my vendor lock-in problem by rate limiting me often enough that I finally tried the alternative. The lesson, unfortunately, appears to apply here too. The next thing you’re dismissing is probably worth trying.
Congratulations, Now What?
Here’s the part the rocket-emoji guys never mention. The technical stuff isn’t the wall. The terminal isn’t the wall. The errors aren’t the wall. The wall of red text isn’t the wall. All of that dissolves the second you start forwarding screenshots to Claude. None of it stays hard for very long.
The real wall is stranger. The moment you realize you can build almost anything, the hard part becomes deciding what to build. The constraint was never my skill. It just moved. It used to be “can I do this?” Now it’s “is this even worth doing?” Which is a much bigger, much weirder question. And nobody warns you it’s coming.
That’s not a reason to stay where you are. It’s the whole reason to start. A blank page where almost anything is possible is the best problem I’ve ever had. I just couldn’t get to it from Level 1. I spent months up there instead, doing the exact thing I’m now begging you to skip. Standing outside the winery explaining the wine.
Go open the terminal. Tonight. I reviewed this entire world from the outside once already, with enormous confidence and almost no firsthand experience. I was embarrassingly wrong. Don’t do what I did.
Stop smelling the wine and drink it.