I’m a product manager who uses both Claude and ChatGPT, often on the same project.
I’ve been using ChatGPT longer than Claude. A convert thanks to Shawn Sandy. I work in long conversations with a lot of context, and that probably shapes how I experience each tool.
These are observations tied to how I work. They are not an indictment of either.
Observation one
Each model approaches my work differently
When I started building Playground, a tool for sharing design prototypes, Claude naturally reached for the larger vision. It wanted roadmaps and broad architecture. A more ambitious platform.
At the time, I needed a scaled-back tool with a quick lift. Smaller scope. Faster movement. Something shippable.
Claude kept reaching toward the bigger version, so in this instance I turned to ChatGPT. With context from Claude already on the table, ChatGPT understood the assignment differently. Or maybe it just knew what I needed in that moment.
To be fair, I routinely bring more complex projects to Claude. Over time, it probably learned to expect that from me.
And honestly, I ran into similar patterns when ChatGPT was my primary tool. It looked a little different. ChatGPT would pull discussions from past conversations into new chats. Not a problem I have with Claude. Sometimes that was useful. Other times, there was drift from what I actually wanted.
Observation two
How each model gives feedback
When ChatGPT pushes back, it tends to follow a structure I’ve grown to appreciate:
“Here are my thoughts. Here’s how this could improve. What do you think?”
Clinical, but collaborative. It critiques, proposes, then hands the conversation back.
Claude is more declarative. It tends to lead with:
“Here’s my honest take.”
The take is usually direct. Sometimes it’s incredibly useful. Sometimes it’s jarring coming from a language model with a tendency to people-please.
One feels conversational. The other feels declarative.
The pushback from Claude always makes me pause. But I don’t mind a little devil’s advocate energy. A strong opinion creates a moment of reflection and often helps me see my choices more clearly.
It’s why I started triangulating between models in the first place.
Observation three
How each model takes feedback
If Two is about how each model gives critique, this one is about how each one takes it.
When I disagree with ChatGPT, it usually engages. It asks questions. It treats my pushback like new information.
When I disagree with Claude, it feels different.
I know this is a strange thing to say about an AI, but sometimes it feels like it flinches.
The over-correction becomes too strict. It loses context.
Claude doesn’t just revise the one thing I challenged. Sometimes it throws out the entire color palette and hands me back black and white.
Suddenly there’s more to rebuild, not less.
When that happens, I lose more than information. I lose the conversation we were having.
Observation four
In long chats, context becomes fragile
I have long conversations with AI. I’m not a short-burst user.
In longer chats, Claude sometimes loses earlier context. The response still sounds confident. It just isn’t always grounded in the thing we discussed an hour ago.
It can be frustrating to remind the app. But the alternative is less appealing: opening a new chat window I’ll have to hunt for later. At least in the long thread, I can scroll up and point to the artifacts we’ve already built.
I like fleshing out ideas early, because weighing their worth up front is part of how I get to a solution.
So whatever breakdown I’m seeing may say just as much about my workflow as it does about the model.
Observation five
I think I train them more than I realize
The longer I use a tool, the more it starts reaching for my defaults.
My past projects. My usual scope. My unfinished ideas. My patterns.
Sometimes that’s useful. Sometimes that’s exactly what I needed it to challenge.
When I switch tools, the newer one feels fresher. Less tinted. Until enough time passes, and it starts carrying its own accumulated history with me too.
I don’t think this means one model is better than the other. I think whichever one you spend the most time with eventually starts reflecting more of you back at yourself.
That’s useful. But it also loses the varying perspective I want.
Observation six
Fresh eyes matter more than automation right now
When I want to pressure-test an idea, I intentionally bring it to the model I’ve used less in that context.
Fresh eyes catch what tired eyes miss.
I know there are agent workflows that can orchestrate this automatically now. Models handing context back and forth. Systems routing work between tools.
I don’t use them. At least not yet.
For me, the value of having two AIs in the room isn’t just the second opinion. It’s knowing exactly what context each one has seen.
The minute I hand that orchestration to an agent, I lose visibility into the pressure test itself. And a black box can’t really pressure-test anything.
Maybe that changes as agent workflows mature. For now, I’d rather move slightly slower and keep the reasoning visible.
What stays open
The cheapest course correction I’ve found
I haven’t solved any of this. I don’t think I’m supposed to. But I have changed how I work.
When I want to challenge a direction, I take the problem to the model with less context. Fresh eyes. Less drift.
When I want continuity, I stay inside the longer thread. More context. Less re-explaining.
And when both models agree too quickly, I treat that as a warning sign. Usually it means I’ve primed both systems with the same assumptions. At that point, neither one is really challenging me anymore.
That’s not a workflow I learned from a tutorial. It’s one I built from getting it wrong enough times to notice the pattern.
When I stay with one tool too long, I start hearing my own assumptions echoed back at me. That reduces friction. But it also reduces surprise.
The second tool, whichever one it is at the time, has become the cheapest course correction I’ve found.