Genspark vs NotebookLM vs Skywork: I Tested 5 AI Slide Tools, Then Built My Own

I think in text. People need visuals. That gap is one of my biggest pain points at work.

I am not a visual person whatsoever. If I made the rules, I would be perfectly happy presenting in markdown: bullet points, tables, well-structured walls of text. That works for me. But apparently this does not register well with people. (Turns out "please just read my outline" is not a presentation style.)

You are always in your own head with your own context. You think people will get you, but they really don't. You need to set the frame, the context first. Visuals are a great hook. They say a picture conveys a thousand words. That's why infographics work so well in capturing people's attention and conveying the message. The visual is the entry point that lets people into your thinking.

So I need to create presentations to communicate effectively. The dream process would be: structure all the text content (with the help of AI, as I figure things out, explore, and debate with myself), then hand it off to an AI to turn it into a beautiful, presentable slide deck. I'd rather outsource the part I don't like.

The Short Answer

Most people who land here searched some version of "Genspark vs NotebookLM", so here it is before the long version:

Genspark respects your structure. NotebookLM understands your meaning. Skywork.ai manages both, and looks better doing it.

That was my conclusion in March 2026, after testing five tools against real work decks. I still stand by it. I also barely use any of them anymore, for a reason I did not see coming when I wrote this post. That part is at the bottom.

What I Was Looking For

I tested five tools: Kimi, Genspark, Canva, NotebookLM, and Skywork.ai. Here's what I was evaluating:

• Context understanding: Does it actually get what I'm trying to say?
• Visual quality: Are the infographics and layouts good enough to communicate?
• Editability: Can I tweak after generation?
• Time savings: Does it actually help, or does it create more work?

That last one is the trap. A tool can save you thirty minutes of design and cost you forty minutes of fixing the order your points arrive in.

The Comparison Table

Skywork.ai

Skywork.ai is extraordinary. The context understanding is almost on point. It gets what I'm trying to say and translates it into slides that actually make sense. The design is more than great; it's genuinely impressive. Editing works through prompting, which fits naturally into how I already work.

The one caveat: detailed editing is possible, but not by default. To make granular changes, you first need to go through an image-to-code conversion step, which takes extra time and can slightly distort images and layout in the process. Worth knowing about, but it never stopped me.

What pushed it to the top is that it gets the best of both worlds. It respects the information structure and order I give it, like Genspark, and the visuals are genuinely great. Neither of the others could deliver both at once.

NotebookLM

NotebookLM (written "Notebook LM" about half the time, including by me) kept all contexts and added a little too much in places, but the infographics are beautiful. The context understanding is the strongest of the five, and it genuinely helps communicate the message.

Two catches. It can't be edited directly, so what it gives you is what you present. And when I bring in a lot of content with a specific talking structure, it tends to shuffle the order. For a lighter presentation that's manageable. When the sequence carries the argument, it's a real problem.

If your comparison is specifically NotebookLM vs Skywork: NotebookLM understands more, Skywork lets you change more. I went with the one I could change.

Genspark

Genspark got the context. It's structured correctly, no message is lost, and it doesn't rearrange things, which matters more than it sounds. It follows the structure you give it. The visual polish isn't there, but for a real stretch of time this was my actual go-to. When I had a lot of information and a specific talking order I needed to preserve, Genspark was the reliable choice.

I'd rather have the right structure with plain visuals than beautiful slides that tell the story in the wrong order.

Kimi

Kimi looks beautiful: the templates are polished and professional. But some messages are lost in the translation from text to slides. When I reviewed the output, I realized it didn't fully capture what I was trying to say. Beautiful slides that say almost what you meant are worse than plain ones that say exactly what you meant, because you don't catch the gap until you're standing in front of people.

Canva

Canva couldn't structure the presentation elements in the right context. The AI didn't understand how to organize my content meaningfully, which made it hard to use. Canva is excellent at a lot of things. Reading my intent out of a wall of structured text was not one of them.

Genspark to Canva, and Other Escape Hatches

A question I see a lot: can you take AI slides somewhere you can actually edit them by hand? Yes. Genspark exports to PPTX, PDF and Google Slides, and there's an official guide for opening those in Canva and Figma. The export route is real and it works.

I never leaned on it, for one boring reason. The moment I open a deck in a design tool, I'm doing the job I was trying to outsource. Dragging text boxes is exactly the part I hate. So I optimized for tools I could correct by talking to them, which is why Skywork's prompt-based editing mattered more to me than anyone's export quality. (I have a whole post about tools that get out of your way. This is the same bias, applied to slides.)

If you like designing and only want the first draft generated, invert that. Genspark to Canva is a perfectly good workflow. It just isn't mine.

Then I Built My Own

Here's the part that makes this post out of date in the best way.

Seven months after writing it, I'm working almost entirely inside Claude Code, with my own context set up: my projects, my naming, my history, my other tools. And the recurring decks broke me. Not the one-off pitch. The weekly one. Same meeting, same shape, same categories, and every single week I was copying the same context into a browser tab that had never heard of any of it before.

Every web tool starts from zero every session. You pay that tax weekly and you stop noticing it. I noticed it.

So I built a slides skill. Then a more general one. Then one for carousels. The mechanism is unglamorous: a data file holds the week's content, a renderer turns it into a deck, and the design tokens live in a file I control. Same input shape every time, same output look every time.

I did not build anything the commercial tools couldn't do. I built the part they couldn't reach, which is everything already sitting on my machine. The skill pulls live state from the tools where the work actually lives, instead of me summarizing it by hand into a text box. It inherits context I've spent a year accumulating. And it chains: the same context that produces the deck also produces the phone-readable summary I read on the way to the meeting, and the carousel if the thing needs to go out publicly. One pass, three artifacts, no re-explaining myself three times.

The output depends on the job. A self-contained HTML deck I present from the browser. A markdown summary for my phone. PDF when it's a carousel. (This is also the first time in my life I have been happy to see a deck as a .html file, which tells you something about my relationship with PowerPoint.)

What was I aiming for visually? Honestly, Skywork's density and Kimi's template polish, which is a strange thing to admit about software you wrote yourself. I knew what I wanted because I'd watched five tools try. But the thing I wanted most wasn't either of those. It was the same look every week without re-deciding. Consistency is the one thing a generative tool structurally cannot promise you, because every generation is a new roll.

I won't pretend this is a universal recommendation. It took real effort, it only pays off on work that repeats, and if I hadn't already been living in this setup daily it would have been a ridiculous thing to build. If you make four decks a year, go pay Skywork. This only made sense because I was already there.

It replaced the tools about 99%. Not 100%, because roughly once a quarter something lands that doesn't fit any shape I've built, and I go back and open Skywork like nothing happened. See the PM co-pilot post for the same pattern in a different domain: once the AI knows your context, the generic version of the tool starts feeling like it's guessing.

Renting Taste

Looking back, the year of testing wasn't wasted time before the real solution. It was the requirement-gathering phase, and I just didn't know I was doing it.

I was renting taste. I couldn't see what a good deck of mine should look like, so I paid other people's tools to show me, over and over, until I'd seen enough versions to recognize my own preferences. Skywork taught me how much information a slide can carry before it collapses. Kimi taught me what polish looks like. NotebookLM taught me that understanding and obedience are different features and I wanted the second one more. Genspark taught me that I will trade beauty for correct order every time.

None of that was knowable in advance. I would have given you a confidently wrong answer if you'd asked me at the start what I wanted.

You rent taste until you know what you actually want. Then you stop paying rent.

Does this solve my original problem? More than before. I can focus on my strength, which is structuring ideas and thinking through content, and let something else handle the visuals. The difference is that the something else now knows what I'm working on before I open my mouth.

And I'm still not sure I'd recommend the ending to anyone. Building your own tool is the kind of thing that sounds clever in a blog post and quietly eats a weekend. The honest version is that I got lucky: my annoyance and my daily working environment happened to line up.

Not perfect. But I stopped searching a while ago, and I didn't expect the thing I stopped at to be something I'd end up writing myself.

(The 1% still goes to Skywork. Old habits.)