There is a specific kind of disappointment that happens when someone switches from ChatGPT to Claude and expects the same experience with a different name on the tab. The prompts they spent months refining do not work quite the same way. The outputs feel different in ways that are hard to articulate. Some things are better, some things are worse, and the whole thing feels slightly off in a way that makes them wonder if they made a mistake switching.
They did not make a mistake. They just brought the wrong mental model with them.
Claude and ChatGPT are not the same product with different branding. They were built by different teams with different philosophies, trained on different data with different objectives, and they have genuinely different behavioral tendencies that show up in ways that matter for how you work with them. Treating them as interchangeable means you get mediocre results from both instead of excellent results from either.
Here is what actually changes when you move from one to the other.
Claude takes instructions more literally than you expect
This is the thing that surprises people most. ChatGPT has developed, over years of product iteration, a tendency to interpret prompts generously. It fills in gaps, makes reasonable assumptions, adds things you did not ask for but probably wanted, and generally tries to give you what it thinks you meant rather than exactly what you said.
Claude does the opposite. Claude reads your prompt and does what it says. Not what it thinks you meant, not what would probably be helpful based on context, what you actually wrote.
This is not a bug. It is a design choice and it has real advantages, especially for applications where predictable, consistent behavior matters. But it means that vague prompts that worked fine with ChatGPT, because ChatGPT was filling in the gaps charitably, produce more literal and sometimes underwhelming results with Claude.
The fix is to be more specific. Not more verbose, more specific. Instead of relying on the model to infer what good looks like, tell it. Instead of assuming the model will add the context it needs, give it. Claude will follow detailed instructions with more fidelity than ChatGPT, which is the upside of the same trait that makes it less forgiving of vague ones.
As Nathan Lambert wrote in his piece on switching from ChatGPT to Claude, getting ChatGPT to behave the way you wanted often required specific override prompts like "no yapping" or elaborate formatting instructions. Claude tends to just do what you ask without needing those workarounds, but that only works if what you asked was actually specific.
Claude pushes back and that is not a defect
One of the most common complaints from people who are used to ChatGPT is that Claude sometimes asks clarifying questions or expresses uncertainty or declines to make assumptions when given an ambiguous request. This feels like friction to people who expect the model to just handle it.
That friction is intentional and, in most cases, useful. When Claude says it is not sure what you want and asks for clarification, it is because your prompt genuinely was ambiguous and ChatGPT was just picking an interpretation without telling you. Sometimes ChatGPT picked the right interpretation. Sometimes it picked the wrong one and you did not realize it until you read the output carefully.
Claude surfaces the ambiguity upfront. That costs you one extra exchange but it saves you from building on top of a misunderstanding, which costs a lot more.
If you find Claude asking too many clarifying questions for your taste, the solution is not to wish it behaved more like ChatGPT. The solution is to write prompts that are specific enough that there is nothing ambiguous to ask about.
The sycophancy difference is real and significant
Simon Willison noted in his breakdown of Claude's system prompt that one of the explicit instructions in Claude's own guidelines is that it never starts a response by saying a question or idea was good, great, fascinating, or any other positive adjective. It skips the flattery and responds directly.
This is not a small thing. ChatGPT has a well-documented tendency toward sycophancy, the tendency to tell you what you want to hear rather than what is accurate. It will validate a flawed plan, praise a mediocre idea, and soften critical feedback to the point of uselessness. This is not malice. It is a training artifact that OpenAI has acknowledged and struggled to fix.
Claude is meaningfully less sycophantic. When you show Claude a business plan with a serious flaw, it will tell you about the flaw. When you show it writing that is not working, it will tell you what is not working rather than finding something nice to say first. For people who are used to ChatGPT's warmer, more validating responses, this can feel cold. For people who actually want useful feedback, it is significantly more valuable.
If you are using AI to get honest assessments, honest edits, or honest second opinions on anything, Claude's lower sycophancy is a real advantage. Take it seriously rather than trying to prompt it into being more agreeable.
Format defaults are different
ChatGPT has a strong default toward markdown formatting. Headers, bullet points, bold text. It structures responses visually even when you did not ask for structure and even when the content would read better as plain prose.
Claude is more contextually adaptive. In a conversational exchange it tends to respond conversationally without reaching for headers and bullets. In a technical or document context it will use structure when it serves the content.
For most users this is an improvement. But if you have downstream systems that expect markdown formatted output and you built that expectation based on ChatGPT's defaults, you may need to add explicit format instructions to your Claude prompts where you did not need them before.
What transfers and what does not
Most of the substantive prompt engineering you learned on ChatGPT transfers to Claude. Giving context, specifying format, using examples, asking for step by step reasoning, all of that works the same way and for the same reasons.
What does not transfer is the meta-layer of managing ChatGPT's quirks specifically. Prompts designed to reduce verbosity, prompts that preemptively tell the model not to add disclaimers, prompts that exist to override specific ChatGPT tendencies rather than to accomplish the actual task. Those prompts are solving problems that Claude mostly does not have, and including them can actually create friction where there was none.
The cleanest approach is to start fresh with your most important prompts rather than porting them directly. Use the same goals but write them from scratch for how Claude actually behaves, with more specificity on the task and less defensive prompting against tendencies that are not there.
The Token Counter on Prompt Toolbox is useful during this process because prompts written for ChatGPT often carry more overhead than Claude needs, and trimming that overhead while rewriting saves both tokens and the cognitive load of instructions the model does not need.
The honest bottom line
Claude is better than ChatGPT for some things and worse for others, and most of that difference comes from behavioral tendencies rather than raw capability. More literal instruction following, less sycophancy, different format defaults, a tendency to surface ambiguity rather than paper over it.
None of those are reasons to prefer one model over the other in the abstract. They are reasons to understand which model's tendencies match your specific use case and to prompt accordingly rather than treating the two as interchangeable and wondering why results feel inconsistent.
