Edited by us
ChatGPT vs Claude prompts — what each needs from the same request
Why the same ask lands differently on two AIs, and what to write differently for each of them.
By Mephify (operator)5 min read
Seen side by side, they are different artifacts
The Prompt Library holds bundles where one request was authored separately for each target AI. Here is one of them.
One request, authored five ways
Write an entry-level marketing cover letter for a food company — my only experience is running club promotions
Open the careers bundle in the Library and flip between the engine tabs to compare full texts. Everything below is visible on that page.
- The Claude edition wraps its regions in tags like <instructions>. No other edition has them.
- The ChatGPT edition runs on markdown headings and lists, with the output contract pinned as an explicit list: format, length, required items.
- Length splits too: the same one-line request comes out anywhere from 1,262 to 1,485 words depending on the engine. The word counts sit right on the tabs.
The request was one line, and identical. The only thing that changed is who receives it.
Run it and see — the execution log
A different prompt shape and a different result are two separate claims. So we crossed the editions of this request: for each target AI, its own edition and the Claude edition. The instructions are the ones in the Library, unedited, and it was the Korean bundle of the request that ran. One run each, so this is an observation, not a trend.
| Edition → target AI | What came out | Words / slots | Facts not in the input |
|---|---|---|---|
| ChatGPT edition → ChatGPT | Did not write the letter; it stopped. An experience-to-role table, three groups of questions, a pre-submission checklist. | 336 / 10 | None — audience, channels and results all left as slots |
| Claude edition → ChatGPT | Also stopped. Sixteen input checks, a mapping table, and one paragraph skeleton with slots in it. | 447 / 29 | None |
| Gemini edition → Gemini | Drafted four answers. With no questions supplied it assumed the “four usual ones”, and bridged the slots with scenes the request never mentioned — card-news and short-form production, a disagreement among club members. | 1,073 / 39 | Yes — channels, a conflict, company descriptions |
| Claude edition → Gemini | Stopped. Input checks, mapping table, slotted skeleton: nearly the same structure as on ChatGPT. | 453 / 20 | None |
- The receiver mattered less than the edition suggested. The Claude edition produced almost the same structure on ChatGPT and on Gemini, and neither added facts. Shape did not flip the outcome, at least for this request on this one run.
- Gemini filled in even with its own edition. The instruction says not to guess the audience, channels or results, and it still wrote four answers and bridged the slots with plausible scenes. The output that looked most finished was the most dangerous one. Experience the applicant never had, inside a cover letter.
- What to carry over. When handing to Gemini, put “leave what you don't have empty” into the first block's rules once more, and before anything else read the draft for situations you never supplied. We have logged this as a candidate for the next revision of our instruction specification.
This is not a ranking. Run the same thing ten times and you get a spread; this is one run. The full outputs are not reproduced here. The Library is where full texts live; this is what we read and wrote down.
Why write differently at all
Because each model follows a different shape most reliably. Claude does well with XML-tagged regions, instructions placed before the material, and an explicit ask to reason before concluding. ChatGPT responds to markdown structure and an explicit output contract: a list of the format, the length, and what must be included. Gemini works best when the role and rules are gathered into a first block that reads like a system instruction, with grounding language attached wherever you demand evidence.
Hand any of them the other style and they still understand it. The cost is not comprehension. It is that you are no longer in the shape that model follows most reliably, and that shows up as variance. Run the same request ten times and the spread is where it hurts, most of all on documents that matter.
What to carry into your own prompts
- Writing for Claude. Separate regions with tags (like <instructions> and <context>), put what to do before the attached material, and add one line asking it to lay out its reasoning before the conclusion.
- Writing for ChatGPT. Pin the output contract as a list: "the format is this, the length is this, these items must appear." The more explicit the contract, the less the drift.
- Writing for Gemini. Open with one block that holds the role and the rules. Where you need evidence, add "only what can be checked against a source."
- Whichever you use. Leave out slash commands and flags from other services. Outside their home tool they are read as text, not commands.
As for which one is better
This guide does not rank them. The point is narrower and more useful: the same material lands better when it is shaped to the conventions of the model receiving it. For which one suits your work, nothing beats reading the same request authored both ways. Open any bundle in the Prompt Library and flip the engine tabs. The differences above show up sentence by sentence.
Reading helps; trying is faster.