The AI Presentation Workflow That Actually Works

July 11, 2026

I was a professional consultant and presentation designer for ten years. I built decks and narratives that helped startups raise over US$100 million.

I had a routine. Learn the company, find the real story, build the argument, copywrite the narrative, design the slides, tweak the layout, improve the visual examples, rewrite copy, rinse and repeat. I spent hours upon hours in slideware (and all kinds of slideware) and I know there are millions of people just like me.

Then everything changed for me. See my division of time:

I love this new division of hours and I'm not going back.

And I'd like to share how I get all this done, saving literally hours every single month (my presentation waves are monthly, for you it might be weekly).

The tools, for me, are Claude Code or Codex plus Obsidian plus Gamma. More precisely: they represent agent + context layer + template + editing system.

If you only need a quick deck and do not care much about editing, this is probably overkill. Use NotebookLM, Gamma or whatever to output something that works.

If you love spending hours on layout and colors and fonts, etc. then this isn't for you either. You're better off sculpting with PPT or even Figma.

It is especially for people who make the same kind of presentation repeatedly: investors, founders, consultants, sales teams, product leaders, anyone who has a recognizable house style and recurring material. If your work has a repeated spine but changing facts, AI should not just make one deck for you. It should let every future deck inherit the structure, taste, and context of the decks that came before it.

That is the part most AI presentation advice misses. This one, for instance. There are tons of surveys like this. They're fine. But I think they miss the point.

The real question is not which tool makes the prettiest deck from a prompt. The real question is: what system makes presentations reliable?

For me, the answer has four parts.

I use Obsidian because my work lives in local files: startup memos, call notes, learning notes, people files, source captures, database rows, screenshots, writing drafts, weekly logs, scripts, workflows. It is backed up, but the working layer is local and agent-readable.

That matters because when I ask Claude Code or Codex to help build a presentation, I am not starting with an empty prompt box. The agent can read my actual work. It can find the latest ARR number, pull the founder note from a call, retrieve the screenshot I filed, compare the company against old memos, and check what I previously concluded.

Nothing else below is possible without this.

This is where Claude Code and Codex matter. I can ask the agent to build a deal-meeting deck on seven companies, and it can read across the corpus before it writes a single slide.

It can decide which company needs a market slide, which needs a business-model slide, which needs a concern slide, and which should probably be cut. It can prepare the substance, pull images, upload assets if needed, and structure the argument before Gamma ever enters the picture.

Before intelligence arrived, this literally couldn't be done, because Zapier/Make style automations were neither smart nor flexible.

I also do not care if it takes 10 minutes or 60. Speed is nice, but reliability matters more. I am perfectly happy telling the agent to assemble a presentation, drive to the office, and arrive with a deck ready to tweak.

Now, this is the part where ppl who know me might say I'm just plugging my own portfolio company, and doubt its value over other slide software.

They'd be wrong. Gamma is genuinely powerful, but not in a way most ppl use it.

For me, Gamma's true power lies in its templates + editing system. Especially templates. Everything changed for me once I started using it.

For years I tried to describe what I wanted from a presentation. Make the title sharper. Use fewer words. Do not write paragraphs. Make it feel like an investor meeting deck. Use the image as evidence, not decoration.

All of that helps, but it is still prose trying to describe taste. And prose is a terrible medium for taste. That's why prompting for taste is a terrible experience.

How? Take a Gamma presentation you like, that you've spent actual time tweaking, and save it as template.

That's it.

Save the deck you actually edited as a Gamma template. The template becomes a reusable workflow asset, not just a one-off prompt.

Next time you want to build a version of the deck, just tell the agent to use that template, and it'll output something 95% there. Seriously.

Like this one, built entirely on an earlier template, getting the card layout and text length exactly like I'd like:

Here's what I think happens: the template captures things I would never remember to specify: title length, image placement, slide density, hierarchy, rhythm, the shape of the argument.

These are all in the "latent space". And AI sees all that, and can codify & repeat those instructions 100x better than I can.

Side note: so many problems in human-AI cowork occur because humans can't see exactly what the AI sees, and AI can't see what humans see.

But Gamma matters for a second reason too: editing. PowerPoint and Google Slides give me too much freedom. I can drag anything anywhere, which also means that when you want to edit content, manual labor goes into keeping everything aligned.

Gamma narrows the space of bad outcomes. For any text block or image, there's maybe 5 spots on the card that you can drag it to. That's it.

It's constrained, and it saves you time.

That is why the final 4% is actually 4%, not another lost evening.

The generated output inherits the template's chart, narration, label, and callout style.

Here is what this looks like in practice.

I capture the work every day. Calls get filed. Memos get updated. Screenshots go into the right folders (I had Claude Code help me build a MacOS feature that captures screenshots and I can quickly enter labels, then later it auto-triage the images with metadata into the right folders.)

That's why when I need a presentation on a company, it takes literally 2–3 minutes.

Claude Code and Codex both know exactly where to find everything: its canonical memo, diligence questions, my conclusions, and image assets.

Gamma template locks in slide structure, headline style, image placement, etc. Information and images go into the right places.

I answer another email, and come back to the deck. Yep, 95% there.

I spent another 5 minutes reading over it, making final tweaks, and I'm ready to go.

For me, as an investor, whether I have conviction in a company, and the 3–5 reasons why, and what risks and blindspots there might be, are literally the most important questions.

To be honest, I wouldn't even need slides to present. They're just visual aid for our colleagues to grasp and feel what I'm talking about.

Most people are still asking which AI tool makes the best deck.

I think the better question is which workflow lets your future decks inherit everything you already know.

Can the system remember your material? Can it preserve your taste? Can it reuse your structure? Can it give you an editable output you trust? Can it move the labor away from formatting and toward thinking?