What I was actually building
I spent the last two weeks building Notavello — a tool that exports AI conversations to color-coded PDFs. Real code, real deadlines, real frustration when things broke. Not a toy project or a tutorial. A product I'm trying to ship.
I used Claude for the majority of it. One day I switched to Copilot to see what I was missing. I switched back by the end of that day.
What Copilot does well — honestly
I want to be fair here because there are real things Copilot does better, or at least differently in ways some people might prefer.
It feels more human. Copilot has a personality. It opens sessions warmly, it called me Mike without being asked, it says things like "let's talk through this together as developers." Some people will find that annoying. Some will find it reassuring. It's a real difference.
No hard usage wall. This is a genuine advantage. Claude has a hard limit — when you hit it, the conversation ends and you get a message telling you you're out. Copilot slows down instead. It gets less responsive, things start taking longer, sometimes it just doesn't reply for a bit. That's frustrating in its own way, but it at least gives you a chance to notice and adjust rather than hitting a wall mid-thought.
It's free. That's not nothing. If you're just starting out and can't justify a subscription, free access to a capable AI coding assistant is genuinely useful.
What Copilot does that drove me up the wall
Here's where I stop being diplomatic.
It cuts off your code silently. Copilot has an output limit — a ceiling on how much code it can generate in one response. When it hits that limit it doesn't tell you. It just stops, presents the truncated output as if it's complete, and waits. You copy the code, paste it, run it, and only then discover that a function is half-written or a closing bracket is missing. I wrote a whole article about this — it's one of the most frustrating things I've encountered in any AI tool.
When caught, it suggested I use Microsoft Word. I'm not joking. When I pushed back on the truncation issue, Copilot suggested that developers dealing with long outputs should break their work into Word documents or PowerPoint slides to manage complexity. I asked what percentage of developers actually do that. It said 0 to 2 percent. So its official workaround for its own limitation is something 98% of its users will never do. That's not a solution, that's deflection.
The confidence loops. Every session Copilot greets you fresh. No memory of what you tried before. So you explain the problem, it says "I see the issue clearly, here's the fix" with complete confidence, you try it, it doesn't work, you explain again, it says "I see the issue clearly now, here's what we missed" — and you do this five, ten, fifteen times. The tone never changes. The confidence never wavers. But neither does the wrongness.
Claude's problems — because I promised honesty
Claude is not innocent here. I've been using it for two weeks and I've hit its walls too.
"I see the problem clearly now." Claude says this. A lot. Especially in long sessions when the context is filling up. It will diagnose a problem with complete confidence, give you a fix, watch it fail, diagnose again with complete confidence, give you another fix — and you'll do this loop multiple times before it actually solves it. It's less frequent than with Copilot in my experience, but it happens, and it's just as maddening.
It degrades in long sessions. As a conversation gets longer — especially past 70 or 80 percent of the context window — Claude gets noticeably worse. It starts forgetting things you established earlier, contradicts itself, misses details it would have caught at the start. It doesn't always tell you this is happening. Sometimes I have to start a fresh conversation and re-explain the context, which costs time.
The hard cutoff is brutal. When you hit the usage limit, Claude ends the conversation. You get a message, you're out, you're notified. No gradual slowdown, no warning five minutes earlier. Just a wall. When you're 90% of the way through solving something difficult, this is genuinely painful.
Why I kept coming back to Claude anyway
After all that — here's the thing. Claude makes fewer errors. Not zero errors, but fewer. And when it gets something right, it tends to stay right. The code works. I'm not going back to fix the same thing three sessions later because it gave me something subtly broken the first time.
The endless loops happen far less often. With Copilot I was going around the same problem ten or fifteen times regularly. With Claude it's more like three or four times on a hard problem, and then it actually resolves. That difference compounds over two weeks of real work. It's the difference between feeling like you're making progress and feeling like you're running in place.
It also handles complexity better. When a problem has five moving parts that interact with each other, Claude tends to hold all of them in mind at once. Copilot more often fixes one thing and breaks another, then fixes that and breaks something else.
The verdict
If you're choosing between the two for serious coding work, I think Claude is worth the subscription. But I'm biased, I told you that upfront, and Copilot Pro might close some of these gaps — I haven't tested it.
What I can tell you is that after two weeks of building something real, Claude is still open in my browser and Copilot isn't.