Somewhere in every conversation about AI decks, the tool audit begins: Gamma, Claude Design, or Figma? Which one finally gets slides right? I have run that audit more times than I want to admit, and every round ends the same way: the tool settles into its own rhythm and forgets my instructions. By the third deck, the output sounds like the tool, and my brand is a guest on its record.

Decks are what my clients ask about most, and the place where I most often watch their AI experiments fail. I know the problem from both ends. Until this spring, every deck I shipped was built by hand, and when a template promised to save me, I bought in and then built by hand anyway, one text box at a time. The decks came out fine, but they consumed my calendar, and every revision meant reopening the file and pushing boxes around by hand. Handing the whole job to the tool went worse. When I asked a tool to just build me a deck, cold, I got back Muzak: my own argument in an elevator arrangement, dumbed down and aimed at nobody in particular. The real problem was system design. The model had no durable way to carry my style, process, or context from one deck to the next.

The failure deserves more sympathy than it gets, because what a cold prompt buys is the Muzak treatment. Muzak took real songs and sanded them down for the background: the voices gone, the edges gone, nothing left that would make anyone look up. Same logic in the chat window: the model has never seen your brand, never sat in one of your meetings, and when you say “make it pop,” it reaches for the statistical average of every deck ever made. A cold prompt hands over everything at once:  your style, your thinking, your content, your process, and the deck comes back smoothed in all four. Switching tools just changes the elevator, but the music stays the same.

So I stopped auditioning tools and built the thing I actually needed. It takes four steps. The first three, a design file, a build engine, and a skill that combines them, get set up once. The fourth, your context, comes fresh with every deck. Staffing matters too: I run the whole process on Fable 5 or Opus 4.8 on the Claude side, or GPT 5.6 Sol High on the OpenAI side. Deck work gets the most senior models available.

Step one is the design file. Google Labs publishes a format called DESIGN.md, a spec for describing a visual identity to a coding agent: colors as named tokens, a type scale with exact weights and sizes, spacing rules, and the things you never should do. I keep six of these files in a folder now. One holds the palette and type for my own presentations, while the others are for clients or my different teaching engagements. If your company has brand guidelines, most of this work is already done. A design file is brand guidelines rewritten for a reader who takes them literally.

Step two is the build engine, where most of my frustrations were solved. A design file governs how the deck looks, and construction is where AI decks actually collapse: text overflowing its box, shapes hanging off the edge of the slide, a preview that looks perfect until PowerPoint opens it. My engine started with Hands-on-deck, a toolkit from Every that installs as a plugin. I used it as a template and rewrote it for my own system. The engine turns coded layouts into editable PowerPoint slides, checks them for geometry errors, and renders them so the agent can inspect its own work.

Step three sits between those two: a custom skill, built so the design file and the engine speak the same language. It acts as the system’s memory, holding my palette, my nine slide templates, the build workflow, and the hard-won design rules I learned by breaking them. Without the skill, I would re-explain my taste and my toolchain at the top of every session. With it, “make the slides” is a complete instruction.

Step four is your context, and unlike the first three, it comes fresh with every deck. It is the step that determines whether the deck contains an argument or merely fills slides. I skipped it in my cold-prompt days, and it showed. A deck can only argue from what it knows. Hand the system a topic, and you get filler; hand it material, and you get an argument. Before any slide exists, the project gets my notes, prior decks, transcripts, and source documents, and where the argument needs facts I don’t hold, the agent researches first and shows me citations I can check.

With all four in place, the building can start, and the outline comes first. This is the rule I defend hardest. Ask for an outline first, then a blueprint: list every slide with its template and actual headline before anything gets built. My first deck on this system went through five complete rebuilds before I settled on the argument and sequence. Outline approval became step zero in the skill, the closest my system comes to admitting fault.

The build itself is the anticlimax, which is the point. Slides compile, the deck renders, and the agent reviews its own output as images. The whole deck is first reviewed as a thumbnail grid to read the sequence the way an audience will, then slide by slide against a checklist. The final item on the checklist is a hard rule on visual clutter: scan every slide, find the most decorative element, and delete it. Human review still earns its keep at the end.

An honest number for the whole system: it gets a deck about eighty percent of the way there. The first draft will never be perfect. Keep the conversation going with your model, make some tweaks, and then go human mode. The last twenty percent stays with me, reading every slide against what I actually meant to say. The division works because the eighty percent it took over, the drafting, the geometry, the consistency, is where the hours went, and the twenty percent it left me is the part I never wanted to hand off.

A week into running the system, I got a look at what the by-hand era never allowed: a hundred and three slides of older material moved into my design system in one day, palette remapped, diagrams recolored, the originals untouched.

My favorite thing about the whole setup, though, is what it refuses to fake. Two of my decks sat finished in every other respect, waiting on ten screenshots of software interfaces that only I could capture, because the skill bans invented visuals the way it bans invented numbers.

If decks are where your AI experiments have gone quiet, start with the design file; it is one afternoon of writing down the taste you already have. And a caveat from the far side of the setup: a deck is one output, and sometimes the wrong one. The same design system and the same content can build a live website instead of slides, an argument that people can visit rather than sit through. Ask me about that one if you’d like to try it.

— Lauren Eve Cantor

thanks for reading!

if someone sent this to you or you haven’t done so yet, please sign up so you never miss an issue.

I’ve also started publishing more frequently on LinkedIn, and you can follow me here

if you’d like to chat further about opportunities or interest in AI, or this newsletter, please feel free to reply.

banner images created with Midjourney.