What Changed When I Went From Chat to Cowork to Claude Code
I was looking for a specification document from about two years ago. I remembered roughly what it was about and approximately what it might be called, which is the normal state of remembering a two-year-old file. I searched for it the ordinary way, typed in the fragment of the name I was most confident about, and got nothing back. I tried two more fragments. Still nothing.
Then I asked Claude Code, in plain language, to find it — the same vague description I would have given a colleague. It didn’t search for the phrase I typed. It broke the description into several plausible terms, ran them as a single search across the filesystem, and returned the file.
That’s a small thing to write down, and it didn’t feel small. What the tool did was the part of the job I normally have to do myself: work out what I actually meant, and only then go and do the thing I asked for.
I have been using Claude in three distinct modes for a while now, and they don’t sit on a smooth curve. They behave more like three different tools that happen to share a name, and the difference between them isn’t how clever the answers are.
Level one: chat
This is where nearly everyone starts, and for an enormous number of questions it’s exactly the right tool. You ask something, you get an answer, you’re done inside a minute. It’s fast, it commits you to nothing, and most days most of what I need from AI is genuinely this — a competent answer to a specific question, immediately.
The ceiling is real, though, and it’s worth stating plainly: chat answers, it doesn’t reach. It has no access to your files, your systems, or anything else you work with. It can describe how to do something in detail and it can’t go and do it. Every limitation people cite when they explain why AI didn’t change much for them is a description of this level.
Most of the public argument about what AI can and can’t do for a business is being conducted by people whose entire experience is here.
Level two: Cowork
Cowork is the first level where it stops being a conversation and starts being something you hand work to. Multi-location access, real workflows, agents that carry out a task rather than explaining how one might be carried out. The underlying model isn’t different. What is different is that it has somewhere to act instead of only somewhere to talk.
That distinction sounds like a technicality until you watch what it does to the kind of request you make. At level one you ask questions you already know how to phrase, because you’re going to act on the answer yourself. At level two you start handing over things you wouldn’t have bothered to ask about, because the asking and the doing were never the expensive parts — the expensive part was you being the bridge between them.
This is the level most people never reach, and I don’t think the reason is difficulty. It’s that nothing at level one suggests there’s anything past level one.
Level three: Claude Code
This is the one that untied my hands, and it’s where most of my actual interest sits.
Not because it’s more articulate — it isn’t, particularly. Because it operates underneath the layer I was working at before. A terminal, real commands, the ability to take something I described badly and decompose it into the several specific operations that would satisfy it, which is what happened with the file search at the top of this article. I didn’t type a better query. I typed a worse one, in ordinary human language, and the translation step between a vague ask and a literal operation stopped being my job.
That translation step is most of what makes a request hard to act on, in any context, with any tool.
Two other things happened at this level that I wouldn’t have predicted. My own phone had roughly two years of photos on it that had never synced anywhere. A repair shop looked at the device, concluded it was probably finished, and was unwilling to attempt a fix because the likely outcome was losing everything on it. I got the data off myself, with Claude Code, over an evening. Separately, I have written custom software for ordinary consumer hardware that made it do considerably more than it was sold to do — nothing exotic, nothing that involved circumventing anything, just my own software running on top of a device that shipped with a narrower idea of its own purpose.
Neither of those is my job, and I mention them for one reason: they’re the moments where the gap between the levels stopped being theoretical for me. The same capability that finds a two-year-old file finds a way to pull data off a device a repair shop had written off.
The serious version of this is the system that runs my own sales pipeline — it drafts outreach, chases deals that have gone quiet, and keeps the record of what happened honest, without sending anything or deciding anything on its own. That’s the thing I would point at if someone asked what level three is actually for. The phone is a story. The pipeline is the work.
The boring one matters too
I also built myself a private weekend digest — a small job that searches for events near me, runs on deliberately cheap models because the task doesn’t need expensive ones, and emails me a formatted summary. Nobody would demo this. It took very little to build and it saves me a recurring half hour of browsing that I never enjoyed.
I bring it up because it’s the honest shape of most of the value at this level. Not a single impressive capability, but a growing pile of small jobs that were previously not worth the effort of doing properly, each one now cheap enough to be worth automating. The interesting ones are rarely the ones that would look good on a screen.
Where the levels actually differ
It’s tempting to describe this as a capability ladder, with each step doing more than the one below it. That isn’t quite what separates them.
What separates them is access. The model at level one is, for practical purposes, the same model as at level three. What changes is how much of your actual situation it can reach — your files, your systems, the record of what your business has already done — and therefore how much of your problem it can address rather than describe. A model with no access can only give you general answers, and general answers are exactly what people mean when they say AI is impressive but not useful to them. They aren’t wrong about what they saw.
This also explains why the gains don’t arrive evenly. Moving from level one to level two changes what you can hand over. Moving from level two to level three changes what you can build. Those are different kinds of benefit, and only the first one is available to someone who hasn’t yet connected the tool to anything they own.
Don’t jump to level three
I’m not making the case that everyone should be running a terminal. Most of what most people need from AI on a given day is a level-one question answered well, and reaching past that when a quick answer would do is its own waste.
The narrower point is about how opinions get formed. A lot of people’s view of what AI can and can’t do was built entirely at level one and then generalised to the whole category. If your best example of trying AI is a chatbot that answered a question reasonably well and nothing since, you have tested the floor and concluded it was the whole building.
If you have been sitting at level one, the useful next step isn’t a bigger ambition. It’s one specific job you do every week that you have never tried handing over, done at level two, with you watching the whole way.