Claude vs GPT: how to prompt each one

Send the same request to Claude and to GPT and you can get two quite different results. Neither is broken. They just read your words differently. Once you know how each family thinks, you can write for the one in front of you, and use the other to check its work.

Why the families differ

Two of the best-known model families are Claude, made by Anthropic, and GPT, made by OpenAI. Each family has bigger and smaller models. Claude Opus and Claude Haiku, for example, are both Claude.

Each company trains its models in its own way, with its own choices about how they should behave. So the families develop different habits, a bit like two capable colleagues trained at different firms. Both are good. They just read the same brief differently.

The main difference comes down to one thing:

• GPT models are literal. They do what you wrote, and little more. If something matters, you have to say it.
• Claude models fill gaps. They work out what you probably meant, and often do more than you asked.

These are tendencies, not rules, and every new version shifts a little. If a model surprises you, notice how, and adjust.

The same request, two results

Imagine sending both families this: “Check the budget spreadsheet.”

A literal model may check it, tell you what it found, and stop there. It did exactly what you asked.

A gap-filling model may check it, fix the errors, tidy the formatting and add a summary sheet. Helpful, unless you did not want the file changed.

Neither is wrong. Each did something reasonable with a short request. The fix is the same for both: say what you want. But what you most need to add is different for each.

Prompting GPT: say everything

A GPT model takes your words at face value. That is a strength. It does not wander off or add things you did not ask for. But it also will not fill in what you left out.

Take this request: “Update the price list with the new supplier costs.” A literal model might update the first sheet only, leave the margins as they were, and not say where it saved the result. Each of those is a fair reading of what you wrote.

Try saying:

Update our price list using the new supplier costs. 1) The price list is “Price list 2026.xlsx” and the new costs are in “Supplier costs March.pdf”, both attached. 2) Update every sheet, not just the first. 3) Keep our margin at 30 percent on every line. 4) Mark each changed row in yellow. 5) Save it as a new file called “Price list 2026 v2.xlsx”. Do all of the above, then report what changed and anything you could not match.

Four habits for literal models

• Number the requirements. A numbered list is easy to follow and easy to check off. It is less likely to skip one.
• Name every file. Say exactly which files and where they are. Do not assume it will go looking.
• Say what done means. “Do all of the above, then report” stops it finishing step one and asking whether to go on.
• Spell out the obvious. Things a person would just know, such as “every sheet” or “keep the same layout”, may need saying.

Sometimes a literal model does part of a job and then asks whether to continue. That is it being careful, not lazy. Tell it plainly: “Yes, do all twelve, without stopping to check with me. Report at the end.”

A quick test for a GPT request: could someone follow it word for word, with no common sense, and still get it right?

Prompting Claude: say what success is and what to leave alone

A Claude model tries to work out what you are really after. That saves typing: a short request often gets a thoughtful, complete answer. The catch is that it may also do things you did not ask for, because it thought you would want them.

Take “Tidy up this proposal.” A gap-filling model might fix the spelling, but also rewrite the pricing section, change the tone and add an executive summary. Some of that may be good. But if the pricing was agreed with the client, the helpful change is a problem.

Try saying:

Tidy up the attached proposal so it is ready to send to Riverside tomorrow. Success means: no spelling or grammar mistakes, consistent headings, and it reads smoothly. Leave the pricing table and the terms and conditions exactly as they are; those are agreed. Do not add new sections. When you are done, list every change you made.

Three habits for gap-filling models

• Say what success looks like. Then let it find the steps. You do not need to number everything, because it fills in the obvious well.
• Say what to leave alone. This is the part it cannot guess. Name what is agreed, fixed or off limits.
• Ask for a list of changes. So you can spot any extras at a glance.

When you read the result, look for anything you did not ask for: a new section, a changed figure, an extra file. If it is useful, keep it. If not, say so, and add that limit next time.

Because it looks for intent, Claude is good at fuzzy jobs. Tell it why you want something and it will often suggest a better way. A quick test for a Claude request: have I said what is off limits?

At a glance

	GPT (literal)	Claude (gap-filling)
A short request gets	A short, narrow answer	A fuller answer, sometimes with extras
What to add	Every requirement, every file, what done means	What success looks like, what to leave alone
Strong at	Reviewing, fact checking, finding things, research	Working with you, running a job start to finish
Watch for	Stopping early, skipping what was not said	Changes you did not ask for

Picking a model and effort

You do not need to think hard about which model to use. A simple order works for almost everything:

• Start with the default. It is chosen to suit most work.
• Move to a bigger model when the work is hard, or the answers feel shallow or keep missing the point.
• Move to a smaller model for quick, simple jobs where speed matters more than depth.

Some models also let you choose how hard they think. More effort is slower and uses more, but helps with tricky numbers, long plans and problems with many moving parts. Rewording a paragraph does not need it.

On a big job you can even mix sizes: a small, fast model to gather and sort, a strong one for the hard thinking.

Try saying:

Use Haiku to pull the totals out of all forty invoices, then use Claude Opus to work out why March is lower than February.

Switching family halfway through

Sometimes the model you are using gets stuck, or runs out of usage. A model from another family often sees the problem differently. In Interweave OS you can switch model in the same chat, even to another family, and the new model is handed the whole conversation and carries on.

When you switch, change your style to suit:

• Moving to GPT: spell out what is still to do, in a numbered list, and say “do all of this, then report”.
• Moving to Claude: say what done looks like, and what it must not touch.

Use another family to check

This is where knowing both families really pays. Models from the same family tend to share blind spots, because they were trained in similar ways. A model from another family is more likely to notice what the first one missed.

The rule is simple. If Claude did the work, have GPT review it. If GPT did the work, have Claude review it.

Try saying:

Have a GPT agent review this report on its own and list any mistakes or weak points. Then tell me where you agree and disagree.

Keep reviews independent. If the second reviewer sees the first one’s opinion, it tends to agree with it, and you lose the point of asking twice.

The same goes for research. Two families searching separately find different sources and reach their own conclusions. Where they agree, you can trust it more. Where they differ is exactly where to look.

Brief each reviewer its own way. Give a GPT reviewer a list of what to check and say “report every problem you find”. Tell a Claude reviewer what is in scope and what to leave alone, or you may get a rewrite you did not ask for.

So which one should you use?

Both. Each is capable on its own, and they shine at different work. Being comfortable with both families also means a limit or a slow day on one never stops yours.

How Interweave OS does it

ChatGPT only offers OpenAI’s models and Claude only offers Anthropic’s. Interweave OS puts every model your organisation connects in one picker, grouped by family, with notes on what each is good for. You can switch family mid-chat without losing the conversation, ask for helpers in any model and family, and press Second opinion under any reply to have a reviewer from another family say where it disagrees.