For as long as I've been doing AI operations work, one question shows up in almost every engagement. It doesn't matter if I'm talking to a marketing team, an executive, a sales org, or an operations group. Someone always asks it.
Can AI just create my presentations for me?
My answer, for most of those four years, has been the same: not yet.
That's not because nobody tried. There have been tools built specifically for this use case, and some of them have real users who swear by them. But they haven't seen the broad adoption you'd expect if they were actually solving the problem, and I think I know why. The tools that exist tend to automate the production work while skipping the thinking work, and it turns out those two things are harder to separate than they look.
Why Presentations Are Harder Than They Appear
A presentation isn't just a document with slides. It's a thinking tool. When you build one, you're not just formatting content. You're making decisions: What's the right order? What deserves a full slide versus a bullet? Where does the audience need to pause and process before moving forward?
A lot of people who ask "can AI build this for me" aren't really asking to be removed from the work entirely. They're asking to get out of the parts that don't require them: hunting for the right layout, spending twenty minutes on alignment, regenerating the same slide five times to get the font right. That's the distinction that matters, and it's the one most AI presentation tools have missed.
What's Working Now
I've spent real time recently building a workflow that lands in the right place. It uses two custom Claude skills: a design system skill and a layout library skill. Together they give Claude everything it needs to generate a fully branded, visually structured PowerPoint file from a plain content brief.
The design system skill defines the visual language of the presentation: a configurable primary color system, a fixed slate palette, typography rules, spacing constants, and a library of shared helper functions. It also includes a layout selection strategy — explicit rules about when to use which layout and how to vary them across a deck so the output doesn't default to the same slide type over and over. Things like "no more than two consecutive content slides" are baked into the instructions.
The layout library skill defines all thirteen slide types as callable functions: cover, section title, standard content, two-column, three-column, split, stats and metrics, icon grid, timeline and process flow, full-bleed image, quote and statement, agenda, and closing. Each function takes a structured options object and produces a complete, formatted slide. When Claude has both skills loaded alongside a content brief, it generates a Node.js script using pptxgenjs that produces a real, editable PowerPoint file.
The output is a presentation that's roughly 80% complete: proper layouts chosen based on content type, consistent brand formatting throughout, and structured content organized across the right slide structures. What's left is judgment work. Swapping in specific images, refining a slide that needs a different structure, tightening the narrative in a few key moments.
That last 20% still requires a person. And I'd argue it should. The parts of a presentation that make it land for a specific audience aren't the parts a system can fully anticipate.
The Part That's Changed How I Work
The more meaningful shift isn't the time savings. It's when and where I can do the work.
Because this workflow runs in the background, I can kick off a presentation from my phone. I give Claude the brief, let it run while I'm in a meeting or in transit, and come back to a near-complete deck. I can have several going at once across different client contexts while I do other work. The constraint that used to exist, that presentations required a laptop, a block of uninterrupted time, and my full attention, is largely gone.
Finishing the Work: Claude Inside PowerPoint
When the generated file is ready, I open it in PowerPoint and pick up where the script left off — using the Claude extension built directly into PowerPoint. This is an important part of the workflow and worth understanding clearly.
The Claude extension inside PowerPoint isn't a chatbot you switch to in another window. It's a native part of the editing environment. You can work slide by slide, ask Claude to refine specific content, pull in research from the web, drag in reference files, and make targeted changes with full context of what's already on the slide. The precision you get there is meaningfully different from working in a general-purpose chat interface.
What this means practically is that the entire process, from initial generation to final polish, stays within an AI-assisted environment. You're not generating a file and then switching to manual editing mode. You're handing off from one AI workflow to another, with the second one giving you the specificity and control the first one couldn't. That continuity matters more than it might sound.
What to Know Before You Try This
This isn't plug-and-play today. Building the skills took real work: testing outputs, learning where Claude needed correction, iterating until the system was reliable. That's a meaningful upfront investment, and not everyone will want to make it.
The technical setup also requires some comfort with Node.js and pptxgenjs. Claude generates the script, but you need to be able to run it and troubleshoot when something is off.
That said, for people doing presentation work at volume, the setup cost pays back quickly. And I expect the friction to come down as these tools mature. The capability is real right now.
Where This Lands
It's not quite the "type in text and get a perfect presentation" answer that people have been asking for. I don't think we're there yet, and I'm not sure we should want to be. The preparation work that goes into a good presentation has value. The thinking, the narrative, the choices about how to visualize an idea: those are worth doing.
What we do have now is a workflow that handles the production burden reliably, runs without your attention, and gets you most of the way there before you sit down to finish the job. For four years my honest answer to that question was "not yet." That answer has changed.