Creating Effective AI Prompts Using ChatGPT and Claude

Why the same prompt gives two different answers

A prompt that works well in ChatGPT usually works in Claude, and the other way round. When the two answers come back different, the wording is rarely the cause. What differs is everything the tool already had before your prompt arrived: an instruction saved months ago, a project, a memory, a tone setting, a file attached in a different way.

A marketing manager pastes last month’s campaign debrief into ChatGPT in the morning and into Claude after lunch, with the same prompt both times. The two recaps don’t match. One is a tight set of bullets and the other runs as paragraphs, and they lead with different results. The prompt didn’t change. What changed was a formatting instruction saved in one tool, a project in the other and a tone setting nobody remembers choosing.

So the useful move is to split the work into two layers instead of hunting for a separate magic prompt for each tool. The task goes in the prompt, written once so that it travels. The habits go into each tool’s own settings, because that’s where the two products store them differently: your role, your reader, your house style and the list of things the tool should not do. This guide runs one ordinary task through both, using what ChatGPT’s and Claude’s own help pages say, all read on October 22, 2026.

One caution before you start. Plans, limits and defaults change often and vary by plan, region and workspace, so treat the numbers below as a snapshot and check them on your own plan. If you want the general craft of building a prompt first, our 4-part prompt formula covers it, and better prompts for better AI results shows how to read a weak answer.

One task, run in both tools

The task is one that comes round every month for many marketing teams: turn a long campaign debrief into a one-page recap for a leadership update. It recurs, you know what good looks like, and you can check the result against the source, which is the same test we use for choosing a first quick win with AI. A marketing manager writes it for a VP, a content marketer for a client, an analyst for a finance partner.

Start with the habits layer. Below is a standing block, filled in for a marketing manager, to paste into each tool’s own settings. It’s short on purpose. ChatGPT’s custom instructions hold 1,500 characters on the Free and Go plans and 5,000 on Plus, Pro, Business, Enterprise and Education,^[1] and a block this size fits all of them.

In Claude, those instructions apply to all of your conversations.^[2] In both tools they’re standing context, which is exactly why the task itself doesn’t belong there.

Now the task layer. This is the prompt you paste into both tools unchanged. Each part sits under a capital-letter label, because separating instructions from material is one habit both vendors recommend. Anthropic’s guide says wrapping each type of content in its own tag reduces misinterpretation,^[3] and OpenAI’s says Markdown headers and XML tags can mark distinct sections and show where reference content begins and ends.^[4] In a chat window, plain labels do the same job without either syntax.

The two guides disagree on one detail: where long material goes. Anthropic’s says to put long documents near the top and the question at the end, while OpenAI’s developer guide says context usually works best near the end of a prompt.^[3] Both pages are written mainly for people building on the APIs, so I’d treat the difference as something to test, not a rule. In the template above, the task and rules come first and the material comes last. If an answer ignores a rule after a long paste, repeat that rule in one line underneath the material and run it again.

Where ChatGPT and Claude keep your context

Much of what you feel between the two tools is a storage question. Where does a standing instruction live, what stays attached to a body of work, and what does the tool carry from one chat to the next? The decision rule I’d use has three tiers: what stays true for every task goes in settings, what stays true for one body of work goes in a project, and what is true only today goes in the prompt.

Projects are the clearest example. A ChatGPT project keeps related chats, files and instructions together, and its instructions apply only inside that project and override your global custom instructions.^[5] A Claude project works on a similar principle: what you upload is used across all chats in the project, and project instructions apply only to chats within it.^[6] The difference you’ll notice is in what sits around them. ChatGPT lets you choose project-only memory, and shared projects use it automatically. Claude gives each project its own memory, and inside a project it uses that project’s memory plus your non-project memory, but not memory from other projects.^[7]

The tone row is the one people trip over. ChatGPT’s Base style and tone setting offers personalities such as Professional, Friendly, Candid and Efficient. It changes how ChatGPT communicates, not what it can do, and for content you request, like an email, the output follows your instructions and context rather than the personality.^[8] Claude’s styles can be switched at any point and apply to new messages, edits and retries in that conversation.^[9]

If a recap sounds off, check the dial before you rewrite the prompt.

For a monthly recap, the table points to a practical set-up. Keep the standing block in settings, make one project per client or campaign family, and put the brand guidelines and last month’s approved recap in the project. The approved recap gives the tool a concrete example of length and tone, which is usually easier to follow than an adjective like ‘concise’.

Which row matters most depends on the seat. A content marketer will spend the most time on the tone row, because voice is the product. An analyst will live in the files section below. A marketing ops lead will care most about projects and memory. And a team lead sharing a project needs to know that what’s in it is visible to everyone with access.

Files and long documents: what to do differently

A common surprise with a long attachment is a quiet one: the answer sounds confident about a chart the tool didn’t actually read. On ChatGPT plans other than Enterprise, PDF retrieval is text-based, so images are discarded and only digital text is extracted.^[10] Claude analyzes both text and visuals in PDFs of 100 pages or fewer, and from 101 to 1,000 pages it processes text only.^[11]

So a content marketer who uploads a results deck, where the headline numbers sit in a chart image, can get a fluent recap built only from the numbers written in text. Nothing in the answer needs to tell you a chart was skipped.

The rule that works in both tools: if a number matters, make sure it exists as text somewhere the tool can read.

Raw size is a smaller worry than it sounds. Claude’s help page allows up to 20 files per chat, and OpenAI’s caps text documents at 2M tokens per file. The limit that bites sooner is attention. In a Claude project with a lot of material, retrieval switches on when project knowledge nears the context window, which means the tool is reading the pieces it judges relevant, not the whole collection. Four habits cover both tools:

• Say what’s attached and what you want from it, in the prompt, not just in the filename
• Ask for a location beside each claim: a page, a section or a slide
• Paste any figure that lives in a chart or screenshot as plain text
• Ask for the sentence in the source behind each number, then check two of them by hand

That last habit is the cheapest one on the list, and it works because you’re checking against the source you already hold. Our anti-hallucination toolkit goes further for documents where a wrong figure would cost you.

Adjusting the answer without rewriting everything

The first recap is rarely the one you send. What matters is whether your next message fixes one thing or reopens everything. Mark the draft against the debrief, not against the other tool’s answer. A second tool is useful as a second reader, but two fluent answers can still share the same error.

When the draft is close but not right, a one-line follow-up usually does more than a rewrite of the whole prompt. A brand manager who types ‘make it punchier’ can lose two figures along the way, because the instruction names a feeling and nothing to keep.

Say what to change and what must stay.

Our guide to prompting like a conversation explains why one change at a time works. What matters here is that the follow-ups are the same in both tools, and so is the check.

An analyst checking a figure and a team lead checking a decision list are doing the same job, against the same source.

What not to paste, and which account you're in

The prompt is rarely where the privacy risk sits. The account is.

A content marketer about to paste a client’s performance export into a chat window glances at the corner of the screen and sees a personal login, not the company workspace. The prompt is identical either way, and the terms it’s covered by aren’t.

On the Free, Plus and Pro plans, OpenAI says it may use your content to train its models unless you turn off Improve the model for everyone under Data controls, and that by default it doesn’t use inputs or outputs from ChatGPT Business, Enterprise, Edu or the API. A Temporary Chat isn’t used for training while it stays temporary.^[12] Claude draws a similar line. Free, Pro and Max accounts have a Model Improvement setting, and when it’s on, chats may be kept in de-identified form for up to five years; incognito chats aren’t used.^[13] For commercial products such as Claude for Work and the API, Anthropic says it doesn’t use inputs or outputs to train models by default.^[14]

Both vendors carve out feedback: if you rate a response, the conversation behind the rating can be used. And a Temporary or incognito chat settles only the training question. Whether your organization allows that data in that tool at all is a separate decision, and your IT or legal policy comes before anything in this guide.

Switching tools without starting over

People switch for ordinary reasons: the company approves one tool, a colleague shares a project in the other, or a plan hits a limit. The mistake is assuming your set-up moves with you. Projects, memories and standing instructions are stored separately by each product, so rebuild the layers on purpose and test.

Step 7 is the one that compounds. If you don’t have a place for working prompts yet, our guide to building a personal prompt vault is a good start. The same goes for a marketing team: one shared task prompt, one agreed account and one named person who keeps both current beat five private versions.

Teams can start that internally. Expert training shortens the learning curve and builds shared capability, so the template becomes how a marketing team works rather than the habit of one enthusiastic person. Training builds the capability, a recurring task like the monthly recap is the workflow design that gives it somewhere to be used, and repeating it each month is the reinforcement that makes it behavior. Adoption is what shows when that sticks. Our corporate training helps teams set it up.

To tell whether it stuck, measure in this order. Logins alone tell you about access, not about the work.

• Use: is the team running the template at all?
• Persistence: are people still running it a month later, without a reminder?
• Impact: do recaps go out with fewer rewrites, or sooner?

Pick one recap you write every month. Paste the standing block into whichever tool your organization has approved, run the portable prompt once, and mark the result against the debrief. If policy allows a second tool, run it there too, and compare both with the source.

Hina Mian, Co-Founder of Future Factors AI

Hina is a marketing strategist with over a decade of hands-on campaign experience across B2B and consumer brands. She writes about using AI to run leaner, sharper marketing without losing the human touch. Future Factors helps professionals and teams build practical AI capability through role-based training, workflow design, and hands-on adoption.

More about Hina →