ChatGPT, Claude and Gemini answer the same prompt in structurally different ways, so set a role and explicit format for all three and adjust for each model's habits. The general shape is on what makes a good prompt.
- GPT leans heavily structured. For casual copy, ask for no headers, no bullet points and a conversational tone.
- Claude learns from examples. Show it a few reference pieces to match tone (Anthropic suggests 3–5 examples); ask for output only to skip preamble.
- Gemini is fast, less distinctive. Use it for speed and research; use GPT or Claude for creative copy.
- Skip step-by-step on reasoning models. Give them the goal and constraints; spell the steps out yourself for GPT-style models, which OpenAI says benefit from very precise instructions.
- Follow each vendor's order. Anthropic puts long documents first; OpenAI puts context near the end of the system message.
The same prompt sent to ChatGPT, Claude, and Gemini will produce three noticeably different responses. Not slightly different — structurally different. Different lengths, different tones, different assumptions about what you want. Understanding those differences is the difference between getting a first draft you can use and getting something you have to rewrite from scratch.
This article covers the practical behavioral differences between the three major LLMs and how to adjust your prompts to get better results from each. It's written from regular use of all three — not from documentation. What the vendors themselves publish is on the LLM cheat sheet.
Default behavior: ChatGPT defaults to structured, comprehensive responses. Given a vague prompt, it will produce a well-organised answer that covers multiple angles — usually more than you asked for. It formats heavily with headers, bullets, and numbered lists. It tends toward thoroughness over conciseness.
What it does well: Long-form structured writing (reports, proposals, outlines), iterative refinement over multiple turns, following complex multi-step instructions, code generation, and tasks where you need a comprehensive starting point you'll edit down.
What to watch for (observed on GPT-4o, not re-tested on GPT-6): its default output is often too structured for conversational copy. "Write a casual Instagram caption" will often produce something that reads like a press release unless you're explicit about format. Add "no headers, no bullet points, conversational tone" to any copy request.
Prompting tips: GPT models respond well to role definitions ("You are a senior copywriter") and explicit format instructions. Use system prompts to set persistent behavior across a session. It handles very long context well — you can paste an entire brand document as context and it will use it throughout the conversation.
Default behavior: Claude defaults to thoughtful, measured responses with a distinct voice. It tends toward prose over bullets and will often push back gently if it thinks your framing is wrong. It's the most willing of the three to disagree with your premise — which is occasionally annoying and frequently useful.
What it does well: Copy with a specific tone or voice, nuanced writing tasks where register matters, analysis that requires judgment rather than just information retrieval, long-context tasks (Claude has one of the largest context windows available), and any task where you want the model to think carefully before answering.
What to watch for: Claude can be verbose in its reasoning. It will often explain why it's about to do something before doing it. If you want just the output with no preamble, add "Do not explain your process. Output only." to the prompt. It also tends to hedge on definitive statements — useful for accuracy, frustrating for confident marketing copy.
Prompting tips: Claude responds exceptionally well to examples. Show it a piece of writing you like and ask it to match the register — it will read the style more accurately than any other model. For copy tasks, give it the brand voice description and a few reference pieces — Anthropic suggests “Include 3–5 examples for best results.”Anthropic, Prompting best practices, read at source 22 Sep 2026. Corrected 22 Sep 2026: this said 1–2 reference pieces. It will produce output that's closer to on-brand than GPT without needing to specify every constraint.
Default behavior: Gemini Flash is optimized for speed and broad competence across task types. It tends toward efficient, direct responses — less verbose than GPT, less stylistically distinctive than Claude. Its strongest differentiator is speed and its integration with Google's information ecosystem.
What it does well: Research tasks that benefit from recent information, quick summaries, multimodal analysis (images, documents, audio), and tasks where you need a fast first draft rather than a polished final output. Gemini Flash in particular is extremely fast — useful for high-volume tasks where you're processing many prompts in sequence.
What to watch for: Gemini's creative writing output tends to be more generic than the other two. For brand copy or content with a specific voice, GPT or Claude will usually produce better first drafts. Use Gemini for speed and research; use the others for quality creative output.
Prompting tips: Gemini responds well to direct, simple instructions. It doesn't need as much scaffolding as GPT for basic tasks. For multimodal tasks (analyzing an image or document), Gemini's integration makes it the fastest option.
Universal Prompt Techniques That Work Across All Three
Role Definition
A role sets tone and focus. Anthropic puts it this way, and no vendor quoted on this site says a role adds ability, so pair “You are a senior copywriter at a luxury fashion brand” with the specifics of the task rather than instead of them.Anthropic, Prompting best practices, as quoted on the LLM cheat sheet: “Setting a role in the system prompt focuses Claude's behavior and tone for your use case.” Corrected 22 Sep 2026: this said all three models “respond significantly better” to a role and that it wins “on every model”, with no source.
Explicit Format Instructions
Left to their own devices, all three models will choose a format. That choice is rarely the one you want. Always specify: word count, structure (prose vs bullets vs headers), length of individual sections, and what not to include ("no introductory sentence explaining what you're about to do").
Negative Constraints
Tell the model what to avoid, not just what to do. OpenAI's own instruction template asks: “What should the model do, and what should the model never do?” "No corporate jargon. No exclamation marks. Never use the word 'innovative'." answers the second half plainly.
Chain-of-Thought — now conditional, not universal
"Think step by step before answering" was reliable advice when models answered in one pass. On a reasoning model it is redundant, and can make output worse. OpenAI: “Since these models perform reasoning internally, prompting them to “think step by step” or “explain your reasoning” is unnecessary.”OpenAI, Reasoning best practices, read at source 16 Sep 2026. An earlier version of this section said the instruction can push reasoning models into "over-deliberation" and still earns its place on fast models; neither is in OpenAI's guides.
OpenAI states the distinction directly: reasoning models "provide better results on tasks with only high-level guidance", while GPT-class models need "very precise instructions." Their analogy — a reasoning model is a senior colleague you hand the goal to; a GPT-class model is a junior colleague you hand the steps to. The two want opposite prompts.
What to do: on a reasoning model, state the goal, the constraints and what "done" looks like, then get out of the way. On a GPT-style model, spell the steps out yourself: OpenAI says those models “benefit from very precise instructions”.OpenAI prompt engineering guide — opened and re-checked 11 Sep 2026: "Generally speaking, reasoning models will provide better results on tasks with only high-level guidance", as against "GPT models, which benefit from very precise instructions"; and "A reasoning model is like a senior co-worker. You can give them a goal to achieve and trust them to work out the details" while "A GPT model is like a junior coworker. They’ll perform best with explicit instructions to create a specific output." developers.openai.com. See also contextual prompting.
What the vendors publish that most guides miss
Anthropic — put long documents first. With inputs above roughly 20,000 tokens, place the documents at the top of the prompt and your question at the bottom. Anthropic reports this can improve response quality by up to 30 percent on complex multi-document inputs. Most people paste the question first out of habit, which is the wrong way round.Anthropic, prompting best practices, read at source 23 Sep 2026: “When working with large documents or data-rich inputs (20k+ tokens), structure your prompt carefully”; “Queries at the end can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs.”
Anthropic — structure with XML tags. Wrap each kind of content in its own tag: <instructions>, <context>, <example>. For examples, 3–5 works best, and they should be varied enough that the model does not latch onto an unintended pattern.
Google — always include examples. Google's guidance is unusually strong here: "We recommend to always include few-shot examples in your prompts," and prompts without them "are likely to be less effective." They also advise leaving temperature and the other sampling parameters at their defaults on Gemini 3.x rather than tuning them — and on 21 July 2026 Google’s API changelog went further, marking temperature, top_p and top_k as deprecated.
OpenAI — order inside the system message. Their documented structure is identity, then instructions, then examples, then context near the end. Note this differs from Anthropic's long-context advice above; the two are not interchangeable, so follow whichever vendor you are prompting.All three opened and re-checked 11 Sep 2026. Anthropic, prompting best practices — long context ("20k+ tokens"): "Put longform data at the top: Place your long documents and inputs near the top of your prompt, above your query, instructions, and examples", and "Queries at the end can improve response quality by up to 30 percent in tests, especially with complex, multidocument inputs"; on examples, "Include 3–5 examples for best results" and make them "Diverse: Cover edge cases and vary enough that Claude doesn’t pick up unintended patterns". platform.claude.com. Google, Gemini prompting strategies: "We recommend to always include few-shot examples in your prompts. Prompts without few-shot examples are likely to be less effective", and on sampling parameters "Although you can modify these parameters, we strongly recommend keeping them at their default values for Gemini 3.x models". ai.google.dev. Google, Gemini API changelog, 21 Jul 2026: "The sampling parameters temperature, top_p and top_k are now deprecated". ai.google.dev. Both Google pages read at source 16 Sep 2026; the prompting-strategies page still carried its keep-at-defaults sentence that day, so the two vendor pages differ in strength, not in direction. OpenAI, prompt engineering guide — the documented developer-message order is Identity, Instructions, Examples, Context. developers.openai.com.
Role, format and exclusions work on all three models. OpenAI calls “think step by step” unnecessary on reasoning models and says GPT-style models want precise instructions, and vendors disagree on prompt order, so follow the vendor you are prompting.
More prompts, when something changes.
Prompts, model guides and workflow notes, sent when there is something new. Free.
SUBSCRIBE FREE ↗