How to Prompt Each AI Differently: A Developer's Guide to Claude, ChatGPT, Gemini, Midjourney, and DALL-E

Most people use the same prompt for every AI model and wonder why results vary wildly. Here's the thing: each model was trained differently, processes input differently, and responds to completely different prompting strategies. Using ChatGPT prompts on Claude is like speaking French to someone who only understands Japanese. You'll get something back, but it won't be what you wanted.

I've spent the last year building tools that interface with all of these models, and the differences in how they respond to prompts are significant. This guide covers what actually works for each one, based on official documentation and extensive testing.

---

Why the Same Prompt Produces Different Results

Every LLM has architectural differences, training data biases, and built-in behaviors that shape how it interprets your input. Claude was trained with XML tags in its training data. ChatGPT responds well to Markdown formatting. Gemini has a massive context window but interprets document position differently than Claude. And image generators like Midjourney and DALL-E use fundamentally different prompt languages.

The key insight here is that prompt optimization isn't one-size-fits-all. The same "expert role + clear instructions + examples" formula needs to be adapted for each model's strengths. Let me break each one down.

---

Claude (Anthropic)

Claude is the precision instrument of the group. It follows instructions literally and does exactly what you ask -- nothing more, nothing less. This is both its greatest strength and the thing that trips people up.

What Makes Claude Different

• XML tags are a superpower. Claude was specifically trained with XML tags in its training data, so using <document>, <example>, <instructions> tags to structure prompts gives significantly better results than with other models.
• Document placement matters. Put long documents at the top of your prompt, with your instructions and questions below. This improves response quality by up to 30% in Anthropic's own testing.
• System prompts for roles. Use the dedicated system parameter to define Claude's persona. Keep task-specific instructions in the user message.
• Extended thinking. For complex reasoning tasks, Claude's extended thinking mode produces substantially better results on math, coding, and analysis.
• Honesty by default. Claude will refuse tasks it considers unsafe or unclear. Give it explicit permission to say "I don't know" and hallucinations drop dramatically.

Claude Prompting Template

System: You are a senior data scientist specializing in customer behavior analysis.

User:
<documents>
  <document>
    <source>Q4 Sales Report</source>
    <content>[your data here]</content>
  </document>
</documents>

<instructions>
Analyze the sales trends in the document above. Extract specific quotes
that support your conclusions before providing analysis.
</instructions>

<output_format>
- Summary (2-3 sentences)
- Key findings (bullet points with quoted evidence)
- Recommendations (numbered list)
</output_format>

Claude Pro Tips

1. Quote extraction reduces hallucinations. Ask Claude to extract word-for-word quotes from documents before analyzing them. This grounds its responses in actual text.
2. Specify what you don't want. Claude's anti-examples are powerful: "Do not include disclaimers. Do not hedge your conclusions."
3. 3-5 few-shot examples dramatically improve consistency. Claude pays very close attention to details in examples, so make sure they align perfectly with what you want.
4. Keep tasks focused. A single focused task with clear boundaries beats trying to accomplish three things in one prompt.

For a deeper dive into Claude-specific strategies, check out our Claude vs ChatGPT comparison and the complete frameworks guide.

---

ChatGPT (OpenAI)

ChatGPT is the Swiss Army knife. It has the broadest feature set -- image generation, voice, custom GPTs, code execution -- and it responds well to a wide range of prompting styles. But getting the best results requires understanding its specific preferences.

What Makes ChatGPT Different

• Markdown formatting is preferred. Use headings (#), bold (**), and structured formatting in your prompts. ChatGPT processes these more reliably than XML tags.
• Delimiters clarify structure. Use triple backticks (```), triple quotes ("""), or XML tags to separate instructions from data.
• Temperature and Top-P matter. For factual tasks: temperature 0.2-0.5. For creative tasks: 0.8-1.0. These aren't just theoretical -- they produce measurably different outputs.
• Reasoning models are different. If you're using o1 or o3, do not use chain-of-thought prompting. These models reason internally, and adding "think step by step" actually degrades performance.
• Structured Outputs guarantee schema adherence. When you need JSON, use Structured Outputs with a JSON Schema instead of just asking for JSON. The model will never hallucinate invalid fields.

ChatGPT Prompting Template

# Role
You are an experienced technical writer for developer documentation.

## Task
Write API documentation for the endpoint described below.

## Reference Data
"""
POST /api/users
Body: { name: string, email: string, role: "admin" | "user" }
Returns: { id: string, created_at: string }
"""

## Requirements
- Include request/response examples
- Note required vs optional fields
- Keep explanations under 2 sentences each
- Use code blocks for all examples

## Output Format
Markdown with H3 headers for each section.

ChatGPT Pro Tips

1. Provide reference text to reduce hallucinations. ChatGPT tends to be more verbose and speculative than Claude. Giving it source material and saying "only use information from the provided text" keeps it grounded.
2. Specify steps for complex tasks. "First, identify X. Second, calculate Y. Third, present Z." ChatGPT follows step-by-step instructions reliably.
3. Use few-shot examples for pattern matching. Include 2-3 input/output examples and ChatGPT will match the pattern precisely.
4. Frequency and presence penalties reduce repetition. If outputs feel repetitive, bump frequency_penalty to 0.3-0.5.
5. For reasoning models (o1/o3): keep it simple. Brief, clear prompts. No chain-of-thought instructions. No few-shot examples. Let the model think on its own.

Our guide on how to write better ChatGPT prompts goes deeper on these techniques.

---

Gemini (Google)

Gemini is the context monster. With a 1M+ token context window, it can process entire codebases, books, or video transcripts in a single prompt. But it handles that context differently than you might expect.

What Makes Gemini Different

• Massive context window. Gemini 2.5 Pro handles over 1 million tokens. This changes how you can structure prompts -- you can feed it entire repositories.
• Document placement is reversed from Claude. Put your data and documents first, then your instructions at the end. Anchor your question with "Based on the information above..." to keep Gemini focused on your data.
• Thinking levels replace chain-of-thought. Gemini 3 introduced a thinking_level parameter (low, medium, high) that controls reasoning depth. If you were previously writing complex CoT prompts, just set thinking_level: "high" and simplify your prompt instead.
• Grounding with Google Search. Gemini can automatically search the web to ground responses in real-time information. Enable the google_search tool for current events or rapidly changing topics.
• XML or Markdown -- pick one. Gemini handles both XML-style tags and Markdown headings for structure, but don't mix them. Pick one format and use it consistently.

Gemini Prompting Template

<context>
[Your documents, data, or code here -- can be very long]
</context>

<instructions>
Based on the information above, analyze the code architecture
and identify potential performance bottlenecks.
</instructions>

<output_format>
1. Architecture overview (2-3 sentences)
2. Performance issues (table: location, severity, recommended fix)
3. Priority actions (numbered list, most impactful first)
</output_format>

Gemini Pro Tips

1. Keep prompts concise with Gemini 3. The newer models respond best to direct, clear instructions. Verbose prompt engineering techniques from older models can actually cause Gemini 3 to over-analyze.
2. Use structured output for reliable JSON. Set response_mime_type: "application/json" with a JSON Schema for guaranteed schema adherence.
3. Context caching saves money. If you're making repeated requests with the same long document, Gemini's context caching gives you a 90% discount on cached tokens.
4. Temperature 1.0 for grounded searches. When using Google Search grounding, the recommended temperature is 1.0 (counterintuitive, but it's what Google recommends).
5. For image generation (Gemini Flash Image): describe scenes, not keywords. Unlike Midjourney's keyword style, Gemini's image generation works best with narrative paragraphs describing spatial relationships, lighting, and actions.

---

Midjourney

Midjourney is an entirely different beast. It's not a language model that generates images -- it's an image model that reads keywords. The prompting philosophy is fundamentally different from any text AI.

What Makes Midjourney Different

• Keyword-driven, not sentence-driven. Comma-separated descriptors outperform narrative paragraphs. "Cyberpunk city, neon lights, rain-slicked streets, wide shot, cinematic" beats a long description.
• Built-in artistic bias. Midjourney adds its own aesthetic flair by default -- dramatic lighting, cinematic composition, stylization. Use --style raw to suppress this.
• Rich parameter system. The -- parameter syntax gives you granular control that no other image AI offers.
• Multi-prompts with weights. The :: separator lets you weight different concepts: space::2 ship::1 gives more emphasis to space.
• No understanding of grammar. Midjourney doesn't parse sentences like an LLM. Word order and specific vocabulary matter more than sentence structure.

Key Parameters

| Parameter | What it does | Range | |-----------|-------------|-------| | --ar | Aspect ratio | e.g., 16:9, 3:2 | | --stylize / --s | Artistic intensity | 0-1000 (default ~100) | | --chaos / --c | Variation in results | 0-100 | | --no | Exclude elements | Keywords | | --style raw | Suppress artistic bias | Flag | | --sref [url] | Style reference image | URL | | --cref [url] | Character consistency | URL | | --iw | Image prompt influence | 0-3 | | --weird | Experimental aesthetics | 0-3000 |

Midjourney Prompting Template

photorealistic portrait of a woman with silver hair,
soft Rembrandt lighting, shallow depth of field,
shot on 85mm f/1.4, moody atmosphere,
dark background --ar 3:4 --s 200 --style raw

Midjourney Pro Tips

1. Specify camera and lens. "Shot on Canon EOS R5, 85mm, f/1.8" produces noticeably more photorealistic results.
2. Use --chaos 50-100 for exploration, then drop to 0 once you find a direction.
3. Permutation prompts for batch testing: a {red, blue, green} car --ar {1:1, 16:9} generates 6 variations automatically.
4. Negative weights are more precise than --no. Use roses::-0.5 instead of --no roses for fine control.
5. Style references (--sref) are game-changing. Paste a URL of the aesthetic you want and Midjourney will match it.

Our Midjourney prompting guide covers these techniques in full detail.

---

DALL-E / GPT-4o Image Generation

OpenAI's image generation has evolved significantly. DALL-E 3 is being deprecated in favor of GPT-4o's native image generation (gpt-image-1), which is built directly into the language model. This changes how prompting works.

What Makes DALL-E / GPT-4o Images Different

• Natural language, not keywords. The opposite of Midjourney. Write descriptive paragraphs as if describing a scene to a human artist.
• Automatic prompt rewriting. ChatGPT rewrites your image prompt before sending it to the model. This adds detail but can lose your original intent. Be very specific to reduce rewriter drift.
• Excellent text rendering. GPT-4o is significantly better at readable text in images than any other model. Spell out tricky names letter by letter: "O-P-E-N-A-I".
• Conversational iteration. You can refine images through multi-turn chat: "Make the lighting warmer," "Move the subject left." No other image AI does this natively.
• No parameter syntax. No --ar, no --stylize. Everything is controlled through natural language or API parameters (size, quality, style).

DALL-E / GPT-4o Prompting Template

A cozy coffee shop interior during autumn. Warm golden light
streams through large arched windows. Floor-to-ceiling bookshelves
line the exposed brick walls. A tabby cat sleeps on a cushioned
windowsill beside a steaming latte. The perspective is from a
comfortable leather armchair, slightly low angle. Soft, diffused
natural light. Photo-realistic style, shot on a wide-angle lens.

DALL-E Pro Tips

1. Describe what you want, not what you don't. Saying "no cars" can actually cause cars to appear. Focus on positive descriptions.
2. Use the natural style parameter for realistic images. vivid produces more dramatic, hyper-real results.
3. Step-by-step for complex layouts. "First, a white background. Then, centered, a vintage typewriter. Above it, floating letters spelling 'CREATE.'"
4. GPT-4o handles 10-20 objects in a scene -- far more than DALL-E 3 could manage (5-8 max).
5. For edits, be explicit about what to preserve. "Change only the sky color to sunset orange. Keep everything else exactly the same."

---

Quick Reference: Model Comparison

| Feature | Claude | ChatGPT | Gemini | Midjourney | DALL-E/GPT-4o | |---------|--------|---------|--------|------------|---------------| | Prompt style | XML tags | Markdown | XML or Markdown | Keywords | Natural language | | Document position | Top (data first) | Flexible | Top (data first) | N/A | N/A | | Instruction position | Bottom | Flexible | Bottom | N/A | Inline | | Best for | Precision, analysis | Versatility | Long context | Creative imagery | Text in images | | Reasoning mode | Extended thinking | o1/o3 (simple prompts) | thinking_level | N/A | N/A | | Structured output | XML/JSON | JSON Schema | JSON Schema | N/A | N/A | | Hallucination control | Quote extraction | Reference text | Google Search grounding | N/A | N/A | | Context window | 200K tokens | 128K tokens | 1M+ tokens | ~60 words effective | ~1000 chars |

---

The Universal Principles

Despite the differences, some prompting principles work across every model:

1. Be specific. Vague prompts produce vague results, regardless of model.
2. Define the output format. Every model performs better when you specify exactly how the response should be structured.
3. Provide examples. Whether it's few-shot examples for text models or style references for image models, showing what you want always beats describing it.
4. Iterate. No model gets it perfect on the first try. Build in a revision step.
5. Use the right model for the job. Claude for precision analysis. ChatGPT for versatile tasks. Gemini for massive context. Midjourney for creative visuals. GPT-4o for text-in-image work.

---

What This Means for Your Workflow

If you're using PromptWizz, we automatically adapt your prompt for whichever target AI you select. The optimization applies the specific techniques each model responds to best -- XML tags for Claude, Markdown structure for ChatGPT, keyword optimization for Midjourney, narrative expansion for DALL-E.

But understanding why these differences exist makes you a better prompt engineer overall. The person who knows that Claude wants documents at the top and Gemini agrees (but ChatGPT doesn't care) is the person who gets consistently better results across every model.

---

Keep Reading

• Claude vs ChatGPT: Which AI Should You Use? -- Detailed feature comparison
• ChatGPT vs Claude for Coding -- Head-to-head for developers
• How to Write Better ChatGPT Prompts -- ChatGPT-specific deep dive
• Complete Guide to Prompting Frameworks -- RISE, RACE, CoT, and more
• Midjourney Prompting Guide -- Full Midjourney parameter reference
• The Anatomy of a Perfect AI Prompt -- Universal prompt structure

---

Using multiple AI models? PromptWizz automatically optimizes your prompt for whichever model you're targeting. Paste any prompt, select your AI, and get a version tuned for that model's specific strengths. Try it free.