I Had a Deck to Build, So I Went Looking for an AI Tool
I’d been invited to present at COMET’s Budapest meetup. The talk was the outcome of an experiment I’d run on StillMind: an AI-readable mirror of the site, built to see if I could win more citations from ChatGPT, Perplexity and Claude. Good material, a real result, a room of people who’d genuinely care about it. I wanted the deck to do it justice.
So I did the modern thing. I went looking for an AI tool to build it for me.
And I had a head start, or so I thought. All the context for the talk already lived in my Claude Code project for StillMind: the research, the notes, the numbers, the framing, sitting in files I’d been working in for weeks. The dream was obvious. Point an AI presentation tool at that context, let it carry the load, walk away with something fast, accurate and good-looking.
That is not what happened.
A Few Hours, and About 60% of What I Wanted
I worked through the obvious ones. Gamma, Beautiful.ai, a couple of the “type a prompt, get a deck” makers. They aren’t bad tools. But every one of them ran into the same wall, and it was always the same wall.
They couldn’t reach my context. Everything that made the talk mine was already structured in my project, and there was no clean way to hand it over. So I was back to copying, pasting and re-explaining, which is the exact work I was trying to avoid.
The AI inside them felt thin. Fine for a first pass on a generic topic, out of its depth the moment the content was specific and technical. It didn’t understand the experiment. It understood “slides about a thing”.
And I kept not liking what it gave me. The preset themes were close enough to be tempting and wrong enough to be annoying. None of them looked like me. None of them looked like my website. I’d tweak, fight the template, tweak again, and land somewhere around 60% of what I actually wanted, then spend more time chasing the last 40% than I’d spent getting there.
A few hours in, with three half-built decks across three tools and not one of them right, I stopped.
The Tool I Wanted Was Already Open
The realisation was almost annoying in how obvious it was. The reason these tools were struggling was that my context and my taste both lived somewhere else: in the project I already had open, in Claude Code.
So why was I trying to drag all of that out to a presentation tool, when I could just ask the tool that was already sitting in it?
I asked Claude Code to build the deck.
Fifteen Minutes to a Deck That Looked Like Me
Fifteen minutes later I had it.
It built the deck in HTML and CSS, and because it had my whole project to draw on, it didn’t approximate my brand. It used it. The same colours as my website. The same fonts. Components that looked like the ones on my actual pages, because it could see how they were built. Nothing to describe, nothing to recreate from a screenshot. It just matched, because the source of truth was right there.
The content was right too, for the same reason. It wasn’t guessing at my experiment from a one-line prompt. It had the notes, the numbers and the framing, so the slides said what I’d actually found, in the order that made sense.
Then it kept being useful. It had everything I’d have gone to a dedicated tool for: a full-screen present mode, presenter notes, all of it. I could export the whole thing as a single file to share with people after the talk, or export it to PDF when I wanted that instead. Flexible in the ways that actually matter on the day.
Iterating was the part that sold me completely. “Tighten this slide.” “Cut that line.” “Let this one breathe more.” Instant. The copy got sharper every pass, faster than I’d manage dragging text boxes around a canvas. By the end it wasn’t 60% of what I wanted. It was 100%, and it had taken less time than any single one of the tools I’d abandoned.
Why the General Tool Beat the Specialist
Here’s the bit worth keeping, because it’s bigger than slides.
The dedicated tools lost on the one thing they were supposed to own. A presentation tool’s whole job is to turn your material into a good-looking deck. But it can only see the material you manage to feed it, and it can only dress that material in themes someone else designed. Claude Code already lived where my material and my design system were. It didn’t need me to bridge a gap, because there was no gap. It’s the same tool I build all my software in now, pointed at a job that has nothing to do with code.
That’s the pattern I keep noticing. The reflex, when you hit a new job, is to go and find the purpose-built app for it. Sometimes that’s right. More and more, though, the better move is to ask whether the general-purpose tool you already work in can reach your context and your craft directly. When it can, it beats the specialist that can’t.
The Honest Catch
I’ll be straight about the cost, because a post with no downside reads like an advert.
There’s really only one: it spends your tokens. If you’re a heavy Claude Code user already brushing up against your budget, building a deck this way will push you closer to it, or over. That’s the whole list. For me it was a non-issue against the hours I’d otherwise have lost, but it’s the honest catch, so there it is.
I’d still open Gamma if I needed a non-technical teammate to co-edit a deck with me live, or if I had no existing context and just wanted something passable in two minutes. That’s its lane. For a deck that had to be mine, though, my brand, my content, my standard, the specialist never stood a chance against the tool that already knew me.
Next time you catch yourself onboarding a new app to do one job, it’s worth a beat to ask: is the answer already open on my other screen?
Key Takeaways
- AI presentation tools struggle on the things that matter most: they can’t reach the context that already lives in your projects, their built-in AI is shallow on specific topics, and their preset themes rarely look like you.
- The fix was to stop exporting my context to a tool and instead ask the tool my context already lived in. Claude Code built an HTML and CSS deck in about fifteen minutes.
- Because it could see my real design system, the deck used my actual colours, fonts and components rather than an approximation, and the content was accurate because it had my notes and numbers.
- It had every feature I’d have gone to a dedicated tool for: full-screen present mode, presenter notes, single-file export for sharing and PDF export. Iterating on the copy was instant.
- The only real cost is tokens. If you’re already near your budget, a deck will nudge you closer to it.
- The wider lesson: before onboarding a purpose-built app for a one-off job, check whether the general-purpose tool you already work in can reach your context and taste directly. When it can, it beats the specialist that can’t.