A client sent me a deck last month with a note: "AI wrote most of it, can you fix the parts that sound like AI wrote it." That's a strange request when you sit with it — she couldn't point to a specific sentence that was wrong, just a general sense that every slide sounded like it came from the same slightly-too-confident narrator. It hadn't. Half the deck was ChatGPT, half was Claude, dropped in without anyone deciding which tool should be doing which job.

That's usually where the "AI wrote my slides" complaint actually comes from — not bad output, but no division of labor.

The "write my slides" prompt that quietly wastes both models

Most people open a chat window, paste a topic, and ask for "10 slides on X." Both ChatGPT and Claude will happily comply, and both will hand back something structurally sound and tonally identical to what the other one would have produced. That's not a knock on either model — it's what happens when you ask a general-purpose writer to also make editorial decisions about pacing, hierarchy, and what belongs on a slide versus in your mouth while presenting.

Slide text isn't an essay with line breaks. A headline that reads well as a sentence often reads badly as a headline — it's too complete, and a complete sentence on a slide invites the audience to read ahead of you instead of listening. That distinction is exactly what gets lost when you ask for "10 slides" in one shot: the model optimizes for a good paragraph, then chops it into bullets after the fact.

What each model is actually good at, once you split the job

In my experience running the same slide-text task through both tools side by side, the difference shows up less in "quality" and more in what kind of thinking each one defaults to. Claude tends to hold onto an argument's structure across a long draft — useful when a slide needs to compress three paragraphs of reasoning into two lines without losing the logic. ChatGPT tends to be faster at generating options — five headline variants, three ways to phrase a data point — which matters when you're iterating on tone rather than building an argument.

This isn't just a personal impression. PowerPoint's newer Copilot tooling now lets you choose which model handles a given task, routing Claude toward more visual, design-heavy work and OpenAI's models toward research-heavy content — which is a fairly direct signal from inside Microsoft's own product about where each one's strengths tend to sit.

One thing people overlook: this isn't a permanent hierarchy. Ask either model to punch up a single headline instead of draft a full section, and the gap between them mostly disappears — the difference is really about task size, not raw writing ability.

A workflow that treats text as five separate passes, not one

What actually works is refusing to let either model do the whole job in one prompt. Instead:

  • Outline pass: ask for the argument structure only — no slide text yet, just what each slide needs to prove and in what order.
  • Headline pass: feed the outline back and ask specifically for headlines under a hard character count. Vague instructions like "keep it short" get ignored constantly; a number doesn't.
  • Bullet pass: ask for parallel grammatical structure explicitly — "start every bullet with a verb" or "keep every line under 8 words." Without that constraint, you'll get a mix of fragments and full sentences that reads uneven even when each line individually is fine.
  • Speaker-notes pass: a separate prompt, because the model that just wrote tight slide copy will otherwise keep writing tight — and speaker notes need the opposite instinct, more room to breathe.
  • Compression pass: paste the whole deck back in and ask what can be cut. This step catches the redundancy that's invisible slide-by-slide but obvious once the deck is read start to finish.

Five passes sounds slower than one prompt. It isn't, in practice — the one-prompt version usually needs three or four rounds of "make it shorter, make it less generic" anyway, just without a clear idea of what's being fixed each round.

The character-limit problem neither model solves for you

Here's the part that surprises people who've only used these tools for full paragraphs: neither ChatGPT nor Claude has a reliable internal sense of how a line of text will actually sit inside a slide's text box. A headline that's technically 40 characters can still wrap awkwardly depending on the template's font size and column width. Ask for "a short headline" and you'll get something that's short in the model's judgment, not short relative to your specific layout.

The fix is unglamorous: give the model the actual constraint from your template — "this title box fits roughly 45 characters on one line before it wraps" — rather than a subjective adjective. Once you do that, both models follow the number closely. Before you do, you're negotiating with a guess.

Where both models drift into the same failure mode

Push either tool for "strategic-sounding" language and you'll get the same handful of words back — synergy, roadmap, ecosystem, leverage. That's not really an AI problem; it's the same instinct a nervous human writer has when a plain sentence "sounds thin." Whether that word earns its place on the slide or is just covering for a decision nobody made yet is something you have to judge case by case — deleting every instance the model gives you isn't automatically an improvement, and neither is leaving all of them in.

The other shared failure is confidence without calibration. Ask either model for a supporting statistic and it will often give you one — sometimes real, sometimes a plausible-sounding number with no traceable source. Treat any AI-generated figure on a slide as a placeholder until you've checked it against something you can point to, not as a finished fact.

A 2026 usage-data roundup from Zebracat put business-professional use of ChatGPT for presentations and reports at 68% in 2025, against 39% among students — which tells you the practice is already mainstream, not that the output is already reliable. Those are two different claims, and conflating them is how unverified numbers end up on a slide in front of an actual investor or client.

Why the writing pass matters more than the deck-generation pass

It's tempting to skip straight to "generate the whole deck" tools and treat the text as a byproduct of the layout. That gets the priority backwards. Attention on a deck front-loads hard onto the first slide or two, which means the handful of lines you spend the most editing effort on should usually be the ones a model drafted first and fastest — not an afterthought once the visual template is locked in. Get the opening argument right in plain text before you let any tool touch the design.

Back to that client's deck: once we split the work — Claude for the argument structure and the dense comparison slides, ChatGPT for headline variants and quick rephrasing — the "sounds like AI" note stopped coming up. Not because either model got smarter overnight, but because nobody was asking one tool to be both the strategist and the copywriter in the same breath anymore. That's usually the actual fix, more than any prompt trick: decide which job you're hiring the model for before you open the chat window.