Goodbye Canva: branded decks without designing them

My managers used to joke about my slides. Not the ideas in them, the slides. Fonts that fought each other, a logo I had stretched to fit, a colour picked off whatever was nearest. I was the person who could carry the room and then lose it on slide four.

So I did what everyone does. Templates. Then a designer, when I could afford one. Then Canva, and a whole evening moving a box two pixels left.

I closed Canva last month. Not because it is a bad tool. Because I stopped doing the design at all, and once you stop, you are paying a subscription to a step you no longer perform.

What replaced it is not a better tool. It is five steps and three skills, read by the AI you already use before it starts. Claude Cowork in my case, though the structure holds elsewhere. This is the whole thing.

The part everyone gets wrong

People assume the leap is that AI can make slides. Every tool can make slides, and they all make the same deck: a title in a font nobody chose, three bullets, a stock photo of a handshake. Competent, and clearly not yours.

The leap is that the machine can be made to know your brand and your argument before it starts. Which means the work is not in the prompt. It is in what the prompt can already assume.

Five parts. Skip one and it shows in the output.

1. The brand guideline, as a file

Not a PDF in a shared drive. A plain text file the agent reads at the start of every session: hex codes, typeface, spacing rules, what the logo may never do, the words you do not use.

A page is enough. Company brand or personal brand, same job.

Without it you get a competent deck in a house style that belongs to nobody. With it, ten decks later, everything still looks like one person made it. Consistency is not taste applied repeatedly. It is a written standard applied automatically.

2. Teach it how you argue

The guideline covers colour. It does not cover structure.

Give it five or six decks you actually made by hand, and one you know was bad. It picks up how much you put on a slide, whether you open on the number or the story, how you close, where you leave silence. Feed it the bad one too, because the contrast teaches more than the good ones do.

This is the step people skip and it is the one that stops the output smelling generic. It is the same move as teaching the AI to write like you, applied to layout instead of sentences.

3. A skill that builds the deck in HTML

HTML, not PowerPoint. You get real layout control, and it renders the same on every machine.

The one I use is UI/UX Pro Max: design tokens, layout systems, presentations with charts. It is the part that stops the output looking like a template someone else chose.

4. A humanizer

The pass that earns the deck its credibility.

AI writing has tells and audiences have learned them. The rule of three in every sentence. The em dash pile-up. “In today’s landscape.” A closing paragraph that summarises what you just read. One of those on a slide and the room quietly decides you did not write it.

So the last pass before I look at anything is Humanizer: thirty-five known patterns, stripped out, the line put back into my own register. Nothing ships without it, this article included.

5. The step I keep forgetting to mention

A beautiful HTML deck is not a deliverable. Nobody wants a link to a file on your laptop. They want a PDF.

And printing HTML to PDF goes wrong in the same four ways every time. The web font falls back to a generic sans, so the type is not your type. Multi-column grids collapse into one column. Text renders heavy or garbled in Preview and Acrobat, because Chrome embedded it as Type 3 fonts. And you get white gaps between pages where a section was pushed onto the next one.

I got tired of fixing those by hand, so I wrote a skill for it: html-to-pdf. It prints through headless Chrome, then rasterises to a font-free PDF, which is the trick that makes the file look the same in every viewer including a phone. One slide per page for decks, flowing A4 for documents.

This is the step that turns the other four into something you can actually send.

How to install a skill

A skill is a folder with a SKILL.md file inside it. You do not need a terminal and you do not need to clone anything.

On the repo page, click the green Code button, choose Download ZIP, then upload that ZIP as a skill in your tool’s settings. That is the whole install. It works in any AI you already use that supports skills, so the same three files run wherever you work rather than being locked to one vendor.

The three:

The first two are other people’s work and both are open. The third is mine, written because I kept hitting the same four printing bugs.

Where I landed, and the honest conditions

Claude Cowork, for now. The skills are portable, so this is a preference rather than a requirement, but Cowork is where the whole chain runs end to end without me stitching the steps together: it reads the brand file, builds the HTML, cleans the copy, and prints the PDF.

Two conditions worth stating plainly. You need a paid plan; none of this works on a free tier. And it is desktop work, because the skills and the print step both run on your machine.

The deeper point is the one I keep making to clients. The output was never the bottleneck. The context was. Once the brand, the voice and the argument live somewhere the machine can read, the deck is assembly, and assembly is cheap. That is the same argument as building the second brain, pointed at slides instead of decisions.

I did not get better at design. I stopped doing it.