Claude 3 Vs Gemini Vs GPT-4: Who Can Make Amazing Powerpoints?

LLMs can reliably generate the *facts* and basic slide structure for a presentation, but they still struggle to produce consistently sleek, polished design—especially when the task requires tight layout control, correct asset usage, and faithful adherence to a specific visual style. That gap matters because many white-collar workflows are dominated by “make the deck” labor: if models can draft content quickly, the remaining bottleneck becomes design quality and iteration, not research.

The exercise starts with a practical test: three leading models—ChatGPT Plus (GPT-4), Claude 3 Opus, and Google Gemini 1.5 Pro—are prompted to generate Python code that programmatically builds a 10-slide deck (two slides per subject) about the “five good emperors of Rome.” The prompts are kept as consistent as possible across models, including the requirement for a creative, advertising-agency style “as if it was designed by an advertising agency,” with slide content framed as a pitch deck and focused on each emperor’s top achievements.

All three models converge on similar implementation choices at the code level, producing functions that generate slides for Nerva, Trajan, Hadrian, Antoninus Pius, and Marcus Aurelius. But the outputs diverge sharply in the details that affect real-world usability. ChatGPT’s code runs and produces slides that look “normal” and largely boring, with content that appears reasonably accurate—though some rendering issues show up when transferred into Google Slides (e.g., bleeding at edges).

Claude’s result is more visually ambitious in concept: it pulls in Paul Smith–inspired color ideas, but the layout remains underwhelming. Elements such as date rulings overlap, and the overall composition lacks the kind of proportional spacing that makes decks look professional.

Gemini’s code is the least stable initially, requiring multiple rounds of self-correction to fix import and other errors. Even after it runs, the deck quality suffers: it repeats or misplaces content, and it appears to misunderstand the requested context—at one point treating “being one of the five good emperors” as an achievement. The most telling design failure is that Gemini attempts to use images that don’t exist, suggesting that without explicit asset constraints (or provided image files), multimodal or media-heavy slide generation can break down.

To probe whether the design problem is specific to slide code, the test shifts to v0.dev by Vercel, which generates a website layout from a prompt. Stylistically, the website output is more interesting, but it still gets factual details wrong (it omits Nerva and includes Marcus Aurelius’ son). The follow-up reprompt—asking it to remake the design using Paul Smith colors—shows how iterative prompting can steer aesthetics, though the color palette may still be constrained.

The overall takeaway is pragmatic: LLMs are getting better at assembling correct content and generating working templates, but producing consistently “sleek” design requires more than raw generation. The remaining work likely involves agents and systems that can iterate on layout, enforce design rules (colors, spacing, typography), and manage assets—turning deck creation into a controlled design loop rather than a one-shot prompt.