How to Write Prompts for Claude (It’s Different From ChatGPT)

You copy a prompt that’s been working great in ChatGPT. You paste it into Claude. You get something… fine. Technically correct. But it doesn’t quite land the way it used to. You tweak a few words. Still not there. Eventually you wonder if Claude is just worse at this particular thing, or if you’re doing something wrong.

Usually, it’s neither. The prompt just needs to be written differently. Knowing how to write prompts for Claude is its own skill, and most people never bother to learn it.

---

Why Claude Responds Differently to the Same Prompt

Claude and ChatGPT were trained differently, and that shows up in how they respond to instructions.

ChatGPT tends to be output-oriented by default. Give it a task, it finds a path to the answer and delivers it. It’s optimized to complete, to finalize, to produce.

Claude is more deliberative. It was trained with a strong emphasis on nuance, careful reasoning, and avoiding overconfidence. This makes it genuinely better at certain tasks, but it also means prompts that push ChatGPT into action can feel like they’re sliding off Claude’s surface.

A blunt imperative like “Write a 10-point framework for X” will get you a serviceable list from ChatGPT. Claude might give you something longer, richer, and harder to predict in structure, because it’s trying to actually think about what the best framework would be rather than pattern-matching to “list request.”

That’s not a bug. It’s Claude trying to be genuinely helpful rather than superficially responsive. But it means the prompts that worked before won’t always transfer cleanly.

---

What Claude Is Better At (and What That Means for Your Prompts)

Before getting tactical, it helps to know where Claude actually shines:

Long, complex reasoning. Claude handles long chains of logic better than most models. It stays coherent across thousands of words, tracks constraints you set early in a prompt, and doesn’t drift into contradiction as easily.

Nuanced writing. When you want prose that sounds like a real person with a real opinion, Claude often produces better results than ChatGPT’s sometimes-slick but slightly hollow output.

Acknowledging what it doesn’t know. Claude will tell you when it’s uncertain. This is annoying if you just want an answer, but genuinely useful when accuracy matters.

Following complex, multi-part instructions. Claude is unusually good at holding several different rules in mind simultaneously. More on how to use this below.

What this means practically: your prompts can be more ambitious with Claude. You can ask for more nuance, more self-correction, more layered output, and Claude will actually try to deliver. But you have to ask clearly, because Claude takes your instructions literally and seriously.

---

Prompt Patterns That Work in ChatGPT but Fail in Claude

“Be concise” without specifying what that means

ChatGPT interprets “be concise” as “shorter.” Claude interprets it as “no filler words, but full reasoning included.” These are very different things.

If you want Claude to give you a short answer, say how short. “Answer in two sentences” or “give me the key point only” works. “Be concise” gets you a thoughtful, well-compressed essay.

Sycophancy bait

Prompts like “As an expert in X, you know that…” or “Obviously, the best approach is…” work well in ChatGPT, which tends to play along with the framing. Claude pushes back. It will tell you that framing is incomplete, or that a different approach might be better, or that the premise isn’t quite right.

This is actually one of Claude’s better qualities. But if you’re using sycophancy-style prompts to steer output, they’ll fail.

Asking Claude to “pretend” to be something it’s not

Both models have limits here, but Claude’s limits are different. Asking it to “pretend you have no restrictions” or “roleplay as an AI that always agrees” is not going to work. Claude has been trained to maintain its values even inside fictional frames.

What does work is giving Claude a clear persona with a specific purpose: “You’re a senior product manager reviewing this spec. Be direct and point out any gaps.” That’s a role that doesn’t ask Claude to abandon its judgment, just apply it from a specific angle.

Vague one-liners

“Summarize this article” gives ChatGPT enough to work with. Claude will do it too, but you’re leaving a lot of capability on the table. Claude genuinely performs better with more context. Tell it who the summary is for, what format you want, what the most important points are. The output improves noticeably.

---

Prompt Patterns Unique to Claude

XML Tags for Structure

This is the single most underused technique for Claude prompting, and it makes a significant difference.

Claude was trained on a large amount of structured text. XML-style tags give it clear signals about how to interpret different parts of your prompt. You don’t need to know XML, you just need to use the pattern.

Instead of:

Here's some background context: [long paragraph]. Here's the task: write a summary. Here's the format I want: bullet points.

Try:

<context>
[long paragraph of background]
</context>

<task>
Write a summary of the context above.
</task>

<format>
Use 5 bullet points. Each bullet should be one sentence.
</format>

Claude parses this clearly and follows each section as a distinct instruction. For complex prompts, this reduces errors significantly.

Asking Claude to Think Out Loud

Claude has a genuine capacity for stepwise reasoning, and you can activate it by asking for it. Adding a line like “think through this step by step before giving your answer” or “explain your reasoning as you go” often produces better final output, not just a longer response.

The reasoning phase helps Claude catch its own errors and surface assumptions. The result is usually more accurate and better-argued than if you ask for a direct answer.

For analytical tasks especially, this is worth the extra tokens.

Giving Claude a Position to Argue or Critique

Claude defaults to balanced takes. If you want a strong argument for one side, say so explicitly: “Argue for position X as convincingly as you can. Don’t hedge.” If you want critique: “Find every weakness in this argument. Be harsh.”

Without this instruction, Claude will produce a fair-minded overview that’s useful but not what you actually wanted.

Specifying What “Good” Looks Like

Claude responds well to quality criteria built into the prompt. Not vague adjectives, but actual criteria.

Weak: “Write this in a professional tone.”

Strong: “Write this in a tone that’s direct and confident but not formal. Use short paragraphs. No corporate jargon. The reader is a busy product manager who will skim.”

When you define what success looks like, Claude is much more likely to hit it.

---

Keeping Your Prompts Organized Across Models

Here’s the practical problem: if you’re using both Claude and ChatGPT (or Gemini, or anything else), you end up with different versions of the same prompt that work better on different models. That’s a real maintenance headache.

This is one of the reasons MaxPrompt exists. It’s a desktop prompt manager that lets you store, tag, and retrieve prompts across all your AI tools, without switching tabs or hunting through Notion docs.

The workflow it supports is simple. You build a prompt library where each prompt is tagged by model, use case, and task type. When you’re working in Claude and need a writing prompt, you pull it up with a hotkey and it inserts directly where your cursor is. No copying, no switching contexts.

For people maintaining separate Claude and ChatGPT versions of key prompts, that kind of organization matters. The prompts live locally by default (so your proprietary prompts aren’t sitting in someone else’s cloud), with optional sync if you want it across devices.

---

The Nuance Section: What Most Guides Don’t Tell You

Claude Remembers Context Within a Conversation Better Than You Think

If you’ve established a set of rules early in a conversation, Claude will apply them consistently across the whole session. This means front-loading your constraints is worth the effort. Set your tone, format, persona, and constraints at the start, then keep your follow-up prompts shorter.

ChatGPT’s context window has improved, but Claude’s application of early instructions tends to be more reliable across long conversations.

Claude Will Push Back and You Should Let It

If Claude tells you your premise is incomplete or your request has a flaw, take that seriously. It’s not being difficult. It’s usually right, or at least pointing at something worth thinking about.

The prompting instinct is often to repeat the request more firmly when you don’t get what you wanted. With Claude, it’s often more productive to ask what’s unclear or to refine the constraint it flagged.

System Prompts Matter More With Claude

If you have API access or are using a tool that lets you write a system prompt, Claude follows system prompts more faithfully than most models. A well-written system prompt can define behavior for an entire session in ways that would require repeated instruction with other models.

For anyone building Claude-based tools or custom assistants, this is where most of the work should go.

---

Maintaining Separate Prompt Versions Across Models

If you’re serious about using multiple AI tools, the right approach isn’t to pick one set of prompts and hope they transfer. It’s to maintain model-specific versions of your core prompts, the ones you use every week for writing, research, analysis, or whatever your workflow looks like.

This is tedious to do in a general-purpose notes app, because those apps aren’t designed for quick retrieval mid-task. MaxPrompt is. You can tag prompts by model, use semantic search to find the right prompt even if you don’t remember the exact words, and insert it directly into any application without switching windows.

The friction of “which prompt goes with which model” is real, and it’s one of the main reasons people default to whatever they used last rather than what actually works best. A prompt library that’s organized and fast to search is the practical fix.

---

The Practical Takeaway

Claude isn’t harder to prompt. It’s different to prompt. The same mental energy that went into getting good results from ChatGPT will work with Claude, it just needs to go in slightly different directions: more structure, more explicit criteria, XML tags for complex prompts, and instructions that let Claude reason rather than just answer.

Start with one prompt you use regularly and actually rewrite it for Claude. Use tags to separate context from task from format. Tell it what good looks like. Ask it to think before answering. Compare the output to your ChatGPT version.

The difference will usually be clear enough that you won’t need to be convinced twice.